The case for agents that watch accounts is well made elsewhere. This is the operating manual we use when we build them: which tasks to hand over and at what level of trust, how each workflow runs on Google, Microsoft and Meta, how to wire it up, and how to roll it out without letting software rewrite your budget at 3 a.m.
What is AI PPC management?
AI PPC management is a layer on top of the ad platforms, not a replacement for them. It reads data from every system that matters, decides what deserves attention and proposes or takes narrowly defined actions.
The difference from older PPC automation is the investigation step. A rule can tell you that cost per acquisition rose 30%. An agent can check whether CPCs, conversion rate, search terms, tracking or the landing page explain it, and hand you a short diagnosis with evidence. That is the core of our AI marketing automation work for paid media teams: the agent saves the hours spent finding out whether there is a decision to make at all.
Native automation vs rules and scripts vs AI agents
These are three layers that stack. Teams get into trouble when they expect one to do another's job, such as asking a script to understand search intent or asking Smart Bidding to notice a broken form.
Three layers of PPC automation
| Criterion | Native platform automation | Rules and scripts | AI agents |
|---|---|---|---|
| Examples | Smart Bidding, Performance Max, broad match, Advantage+ campaigns | Automated rules, Google Ads scripts, Microsoft Advertising scripts | An LLM-based agent with tools for ad APIs, GA4, CRM and site checks |
| What it optimizes | Bids, placements and audiences inside one platform | Exactly the condition you wrote | Attention: what changed, why, what to do next |
| Data it sees | That platform's signals plus the conversions you send | One account's data, plus files you fetch | Ad platforms, analytics, CRM stages, site health, change history |
| Handles ambiguity | Inside its objective only | No | Yes: classifies intent, reads change logs, explains causes |
| Main risk | Optimizes hard toward a flawed conversion signal | Silent breakage when accounts change | Confident, wrong advice if data or guardrails are weak |
| Best use | Bidding and delivery at scale | Deterministic checks and bulk changes | Monitoring, diagnosis, triage, drafting, reporting |
Native platform automation
- Examples
- Smart Bidding, Performance Max, broad match, Advantage+ campaigns
- What it optimizes
- Bids, placements and audiences inside one platform
- Data it sees
- That platform's signals plus the conversions you send
- Handles ambiguity
- Inside its objective only
- Main risk
- Optimizes hard toward a flawed conversion signal
- Best use
- Bidding and delivery at scale
Rules and scripts
- Examples
- Automated rules, Google Ads scripts, Microsoft Advertising scripts
- What it optimizes
- Exactly the condition you wrote
- Data it sees
- One account's data, plus files you fetch
- Handles ambiguity
- No
- Main risk
- Silent breakage when accounts change
- Best use
- Deterministic checks and bulk changes
AI agents
- Examples
- An LLM-based agent with tools for ad APIs, GA4, CRM and site checks
- What it optimizes
- Attention: what changed, why, what to do next
- Data it sees
- Ad platforms, analytics, CRM stages, site health, change history
- Handles ambiguity
- Yes: classifies intent, reads change logs, explains causes
- Main risk
- Confident, wrong advice if data or guardrails are weak
- Best use
- Monitoring, diagnosis, triage, drafting, reporting
My rule of thumb: let the platforms bid, let scripts and code do the arithmetic, and let the agent do the reading, correlating and writing that used to eat a paid media manager's mornings.
Which PPC tasks should AI handle, and which stay human?
Decide the automation level per task, not per tool. We use three levels: the agent acts or reports on its own; the agent drafts and a person approves; a person decides and the agent supplies evidence.
Fully automated: the agent checks and reports
| Task | Google Ads | Microsoft Ads | Meta Ads |
|---|---|---|---|
| Daily digest of what needs attention | Yes | Yes | Yes |
| Spend pacing and month-end forecast | Campaign and account | Including shared budgets | Campaign or ad set budgets |
| Anomaly alerts with first-pass diagnosis | Yes | Yes | Yes |
| URL, redirect, speed and tag checks | Every active final URL | Same, plus imported tracking templates | Ad destinations and lead forms |
| Disapprovals and policy status | Ads and assets | Editorial status | Ad review and account quality |
| Waste report | Search terms, Performance Max search insights | Search query report, Google import checks | Placement, audience and creative breakdowns |
AI-assisted: the agent drafts, a person approves
| Task | Google Ads | Microsoft Ads | Meta Ads |
|---|---|---|---|
| Negatives and exclusions | Ad group, campaign, account and shared lists | Negative lists or Google sync | Audience, placement and brand safety exclusions |
| Budget changes and reallocation | Campaign and shared budgets | Same | Campaign or ad set budgets |
| Bid strategy targets | Target CPA, target ROAS | Same | Cost cap, bid cap, ROAS goal |
| New ad copy and assets | Responsive search ads, Performance Max assets | Imported or native ads | Primary text, headlines, creative variations |
| Pauses | Keywords, ads, ad groups | Same | Ads, ad sets |
| Offline conversion values | Values per CRM stage | Offline conversions | Conversions API events per stage |
| Client summaries and emails | Yes | Yes | Yes |
Human-only: the agent supplies evidence at most
- Strategy, offers, positioning and new markets.
- Total budget, CPA or ROAS targets, and the volume-vs-efficiency trade-off.
- What counts as a conversion and how stages are valued.
- New campaign architecture and launches.
- Claims in regulated categories, and decisions about Meta's Special Ad Categories (such as housing, employment, and credit or financial products).
- Account access, billing, policy appeals and suspensions.
- Hard conversations with a client or a sales team.
Some automated checks can graduate to actions through a written playbook. Pausing an ad whose final URL returned a 404 on two consecutive checks is a good candidate. Changing a budget almost never is.
How should an agent handle search terms and negative keywords?
The agent classifies search terms by intent, costs them, protects what converts and proposes negatives at the right match type and level for a person to approve. It should not exclude every term without conversions, because low-volume terms often convert later or offline.
The workflow we build:
- Pull 30 to 90 days of terms with cost, conversions and value, including Performance Max search term insights.
- Break terms into n-grams. Words like "jobs", "salary", "free" or a competitor's name reveal waste no single query shows.
- Classify intent with the model: job seeker, DIY, free, wrong service, wrong location, competitor, informational, relevant. A model does this well; a regular expression does not.
- Apply cost thresholds in code, such as spend above a set share of target CPA with no conversions, given the term's clicks.
- Check the protect list: brand terms, core services and anything that produced a CRM lead, even without a platform conversion.
- Propose the negative with match type, level (ad group, campaign, account or shared list) and the cost it would have saved.
- Check conflicts before applying. A negative that blocks your own active keyword is the most common self-inflicted wound here.
- Re-check after two weeks whether impressions on good terms dropped.
A hypothetical case: a four-location HVAC company in the Dallas–Fort Worth area runs Search and Performance Max, imports to Microsoft Ads and runs Meta lead ads. The agent finds "hvac technician jobs", "hvac school near me" and "hvac certification cost", groups them as job seekers, and proposes "jobs", "school" and "certification" as phrase-match negatives on a shared list. It marks "ac repair cost" as protected, because those queries produced booked jobs in the CRM.
Meta has no search terms, so the same workflow reviews placement, audience and creative breakdowns and lead form answers, and proposes exclusions or form changes.
How do you automate budget pacing and reallocation?
Automate pacing as math and reallocation as a proposal. The agent forecasts where each budget will land, explains why it is off track and recommends capped moves; a person approves them.
Pacing needs two numbers: expected spend to date (monthly budget × days elapsed ÷ days in the month, adjusted for your weekday pattern) and projected month-end spend (spend to date + the last seven days' average daily spend × days remaining). Then apply each platform's mechanics. Google allows a campaign to spend up to twice its average daily budget on a given day, while not charging more than 30.4 times that average over a month (Google Ads Help), so the agent should alert on the trend, not the day. Meta budgets sit at the campaign level (Advantage campaign budget) or the ad set level, and Microsoft Ads adds shared budgets.
A hypothetical example: the HVAC company plans $18,000 a month on Google. On day 12 of a 30-day month it has spent $8,900 against an expected $7,200; at its recent $820 a day it would finish near $23,700. The alert does not stop at the number. The agent finds in the change history that the Performance Max budget was raised on the 5th during a heat wave and never reset, and proposes the reset for approval.
Reallocation is harder, because the right question is about marginal returns, not averages: a campaign with a great average CPA may be saturated. The logic we encode:
- Receivers: campaigns losing impression share to budget (not to rank) with recent CPA or ROAS inside target.
- Donors: campaigns not limited by budget whose extra spend buys conversions above target cost, checked against budget and bid simulators where available.
- Caps: a maximum percentage per move per week, and a cooling-off period.
- Learning periods: Meta's guidance is that an ad set generally needs about 50 optimization events in the week after a significant edit to exit the learning phase (Meta Business Help Center), and big budget jumps count. Smart Bidding also needs time.
- Conversion lag: never judge a move on the last few days, since conversions arrive late.
- Data health first: no money moves into a campaign with an open tracking or landing page alert.
How does an agent detect anomalies without flooding you with alerts?
Combine statistical detection in code with diagnosis by the model, and alert only above a materiality threshold. An agent that sends 40 alerts a day gets muted within a week, and then it misses the one that mattered.
| Method | How it works | Good for | Weak spot |
|---|---|---|---|
| Static threshold | Alert when CPA passes a set value | Hard limits, small accounts | Ignores seasonality |
| Same-weekday baseline | Compare with the same weekday over 4 to 8 weeks | Lead gen with weekday rhythm | Holidays, promotions |
| Robust z-score | Distance from the median, scaled by median absolute deviation | Noisy metrics with outliers | Needs history per segment |
| Forecast band | A seasonal model predicts a range | Larger accounts, e-commerce | Setup and upkeep |
| Paired rules | Spend up with conversions flat; clicks normal, conversions zero | Broken forms and tags | Tuning per conversion action |
When something trips, the agent walks a diagnosis tree. CPA up: is it CPC or conversion rate? Which campaign, device, location or day? What changed in the change history? Are conversions still recording in GA4 and the CRM? Is the form working?
Every alert gets a severity, a minimum spend, deduplication so one root cause produces one alert, and a snooze. Three good alerts a week beat thirty mediocre ones a day.
Why does conversion tracking need CRM reconciliation?
Because ad platforms count conversions and businesses count customers, and the two drift apart. Duplicate form fires, spam, untracked calls, attribution windows and consent all pull platform numbers away from what sales sees. An agent can reconcile them every day:
- Store click identifiers with each lead: GCLID, GBRAID or WBRAID for Google, MSCLKID for Microsoft, Meta's click and browser identifiers, plus UTM parameters.
- Compare by campaign: platform conversions vs CRM leads vs qualified leads vs opportunities vs closed deals, for the same dates.
- Explain the gaps: duplicates, test leads, spam, leads missing a click ID (often redirects or forms stripping parameters), or a conversion action counting page views.
- Send qualified outcomes back: offline conversion imports and enhanced conversions for leads on Google (matching hashed first-party data such as email), offline conversions on Microsoft, and the Conversions API on Meta, deduplicated against browser events with a shared event ID.
- Value the stages so Smart Bidding and Meta's delivery optimize toward revenue, not form fills.
The agent prepares the reconciliation and the upload; a person decides what counts as a conversion and what each stage is worth, because that choice shapes every automated bid. Sending customer data to ad platforms also has consent and privacy implications, so review the data flow and platform terms with counsel first.
How do you monitor landing pages and tracking automatically?
Test the pages your ads actually send traffic to, on a schedule, the way a customer would. Pull every active final URL from the ad APIs rather than from a list someone maintains, then check:
- Status and redirects: a 200 response, no redirect chains, tracking parameters intact.
- Speed: a load-time budget per page template.
- Tags: the Google tag, the Meta pixel and call tracking firing on the page and the thank-you step.
- Forms: a scheduled test submission, marked as a test lead, that must reach the CRM and stay out of reporting.
- Phone numbers: dynamic numbers swapping and routing correctly.
- Policy signals: "destination not working" disapprovals, which often arrive hours after the page broke.
When a check fails, the agent correlates it with performance: a form broken by a site release shows up as normal clicks and zero conversions. Here a pre-approved playbook (pause the affected ads, notify the site owner, open a ticket) can run without waiting for someone to read Slack.
Can AI write ad copy that stays on brand?
Yes, if the brand rules live outside the model and every draft is checked in code before a person sees it. The model writes; the rules file, the validator and the approver decide.
The rules file holds tone, approved and forbidden words, claims that need substantiation ("best", "guaranteed", "#1"), required disclaimers, competitor policy, services, locations and currently valid offers. The validator enforces platform limits mechanically: Google's responsive search ads allow up to 15 headlines of 30 characters and 4 descriptions of 90 characters. Never trust a model to count characters.
- Generate from evidence: top-converting search terms, CRM objections, reviews and the landing page, not just a product name.
- One hypothesis per variant: "same-day service" vs "upfront pricing" is a test; ten random rewrites are noise.
- Review platform-generated assets: Google's automatically created assets and Meta's Advantage+ creative enhancements can generate or alter text and images. Decide per account which are on, and have the agent list what the platform generated.
- Respect special categories: in regulated industries and Meta's Special Ad Categories, a person signs off on every line.
- Keep the trail: who approved which text, when, from which draft.
How should AI handle PPC reporting and client communication?
Let code produce every number and the model write the narrative around verified numbers, then check the narrative against the data before anyone sends it.
Our weekly report answers four questions: what happened, why, what we did and what we need from you. The agent pulls figures, compares them with targets and the prior period, writes the explanation and attaches evidence. A verification step extracts every number from the draft and checks it against the source table; any mismatch blocks the report. For agencies, the same pipeline drafts client emails in each client's vocabulary ("booked jobs", not "conversions"), but a person sends them. Clients forgive a late report much faster than a wrong one written with confidence.
What can an agent actually see inside Performance Max and Advantage+?
Less than in classic Search or manual campaigns. Both campaign types move placement, audience and creative decisions inside the platform's model, and reporting is partial by design.
| Performance Max (Google) | Advantage+ campaigns (Meta) | |
|---|---|---|
| What you can see | Asset group and asset reporting, search term insights, and the channel and placement views Google has been adding | Campaign, ad set and ad results, breakdowns by age, placement and region, creative results |
| What you can steer | Conversion goals and values, budgets, targets, assets, audience signals, feeds, brand exclusions, negative keywords | Conversion events and values, budget, cost or ROAS goals, creatives, existing customer definitions, brand safety controls |
| What stays inside the model | Spend allocation across inventory, query-level bidding, most audience logic | Audience selection, most placement and creative allocation |
| What the agent watches | Feed errors, asset approvals, conversion health, search category drift, pacing | Event quality, creative fatigue, frequency, CRM lead quality |
So the agent guards the inputs (signals, values, feeds, assets, exclusions) and judges outcomes against the CRM, since platform reporting grades its own homework. For larger budgets, plan periodic incrementality checks such as geographic holdouts or platform lift studies, designed by a person and monitored by the agent.
What does a PPC agent's architecture look like?
A PPC agent is a scheduled pipeline with a reasoning step in the middle, not a chatbot with the keys to your ad account. Data comes in, code computes the checks, the model investigates and writes, and every change passes an approval queue before an API call executes it.
-
Scheduler starts the run
-
Data layer pulls ads, GA4 and CRM
-
Code computes pacing and anomalies
-
Agent investigates and drafts actions
-
Person approves, API applies and logs
The layers are those of any well-built AI system. The scene below uses a different example, a clinic's patient assistant, but the stack maps one to one:
Illustrative example Fictional business, people and data. Not a real client, conversation or result.
| Layer | In a PPC agent |
|---|---|
| Data | Nightly and intraday pulls from ad APIs, GA4 and the CRM into one database, with shared metric definitions |
| Model | A language model with task prompts for triage, classification, diagnosis and drafting |
| Tools | Read tools (reports, change history, URL checks), propose tools (write to the queue), a few capped apply tools |
| Guardrails | Spend and change limits, protect lists, brand rules, approval rules, a kill switch |
| Evaluation | A test set of past incidents, alert precision tracking, weekly review |
Three design decisions matter most. Separate read, propose and apply, so most tools only read. Log everything (what the agent saw, its proposal, the approver, the API response) so any change can be traced and reversed. Deliver where people work: a digest in Slack or email and one-click approvals. For the general method, see how to build an AI agent.
Which APIs and tools connect an agent to your ad accounts?
Each platform has an official API, and most setups add a database and a workflow tool. Choose by what the agent must do: read, write, run on a schedule or reach several systems at once.
| Integration | What it gives the agent | Watch out for |
|---|---|---|
| Google Ads API | Reporting, change history, budgets, bids, keywords, negatives, ads, conversion uploads | A developer token approved by Google, OAuth, version upgrades |
| Google Ads scripts | JavaScript inside Google Ads, schedulable as often as hourly | Run-time limits; no CRM access unless you build it |
| Microsoft Advertising API and scripts | Reporting, bulk changes, scripts similar to Google's | Google imports can overwrite Microsoft-side edits |
| Meta Marketing API and Conversions API | Campaign management, insights, audiences, server-side events | App permissions and rate limits |
| GA4 Data API and BigQuery export | Sessions and events outside ad platform attribution | Event-level data needs BigQuery |
| CRM APIs (HubSpot, Salesforce, Pipedrive) | Lead stages, values, click IDs, outcomes | Field hygiene decides everything |
| n8n or Make | Scheduling, API calls, branching, approvals | You maintain the workflows; see n8n vs Make vs Zapier |
| MCP servers | Controlled, standard tool access for AI models | Scope credentials and log calls |
The Model Context Protocol puts the ad APIs, the CRM and your database behind a defined set of tools any compatible agent can call, with permissions, logging and limits in one place. Before connecting a published MCP server for an ad platform, check whether it is read-only, which credentials it holds and what it logs. When none fits, we do MCP server development to expose exactly the tools an agent should have (what MCP is covers the basics). Connecting older CRMs and call tracking is usually the longest part of a project, and the core of our AI integration work.
Should you build or buy PPC automation tools?
Buy when your needs match what a PPC platform already does; build when the value sits in your CRM data, your rules and your way of working. Many teams end up with both.
Build vs buy for AI PPC management
| Criterion | Native platform tools | PPC platforms (Optmyzr- or Adalysis-type) | Workflow tool + LLM (n8n, Make) | Custom AI agent |
|---|---|---|---|---|
| Time to value | Immediate | Days | Weeks | Weeks to a few months |
| CRM, site and GA4 data | No | Partial, via integrations | Yes, through the APIs you connect | Yes, designed around it |
| Custom logic and brand rules | No | Templates and rule builders | Limited by the visual builder | Full |
| Approvals, logs, rollback | Change history | Built in, varies by vendor | You build it | Designed to your standard |
| Maintenance | None | Vendor | You | You or your partner |
| Cost structure | Included | Subscription, often tied to spend or accounts | Platform fee, model usage, build time | Build investment, hosting, model usage |
| Best for | Every account, as the base layer | Proven audits, alerts and ad testing | Small teams with a technical owner | Multi-account, multi-channel, CRM-driven goals |
Native platform tools
- Time to value
- Immediate
- CRM, site and GA4 data
- No
- Custom logic and brand rules
- No
- Approvals, logs, rollback
- Change history
- Maintenance
- None
- Cost structure
- Included
- Best for
- Every account, as the base layer
PPC platforms (Optmyzr- or Adalysis-type)
- Time to value
- Days
- CRM, site and GA4 data
- Partial, via integrations
- Custom logic and brand rules
- Templates and rule builders
- Approvals, logs, rollback
- Built in, varies by vendor
- Maintenance
- Vendor
- Cost structure
- Subscription, often tied to spend or accounts
- Best for
- Proven audits, alerts and ad testing
Workflow tool + LLM (n8n, Make)
- Time to value
- Weeks
- CRM, site and GA4 data
- Yes, through the APIs you connect
- Custom logic and brand rules
- Limited by the visual builder
- Approvals, logs, rollback
- You build it
- Maintenance
- You
- Cost structure
- Platform fee, model usage, build time
- Best for
- Small teams with a technical owner
Custom AI agent
- Time to value
- Weeks to a few months
- CRM, site and GA4 data
- Yes, designed around it
- Custom logic and brand rules
- Full
- Approvals, logs, rollback
- Designed to your standard
- Maintenance
- You or your partner
- Cost structure
- Build investment, hosting, model usage
- Best for
- Multi-account, multi-channel, CRM-driven goals
Two honest caveats. "AI" in a PPC tool can mean anything from a chat window over reports to autonomous changes, so ask exactly which actions it takes without approval and what it logs. And custom agents are not free to run: someone owns API version changes, model updates and the test set. Our breakdown of AI automation cost explains what drives build and running costs. When the job is mostly moving data between tools, start with workflow automation and add the agent where judgment is needed.
What guardrails does a PPC agent need?
Tie each risk to a specific control. This is the minimum before any write access:
PPC agent governance
If your organization has no rules yet for how staff and systems use AI, write them first; our guide to an AI governance policy is a practical start.
A 30/60/90-day rollout plan
Roll out in three phases (observe, assist, then act within limits), each ending with a review that decides whether the next one starts.
From read-only to approved actions
-
Step 1
Days 1–15: Connect and baseline
Read-only access to ad accounts, GA4 and the CRM; record hours spent on checks and reports and how long recent incidents took to find.
-
Step 2
Days 16–30: Digest and alerts
Pacing, anomaly and landing page checks delivered daily; tune thresholds until most alerts lead to an action or a conscious "no action."
-
Step 3
Days 31–45: Search terms and reconciliation
Weekly n-gram reports with proposed negatives and a daily platform-vs-CRM reconciliation; nothing applied without approval.
-
Step 4
Days 46–60: Approval queue and reporting
Budget, pause and copy proposals go through one-click approval; the weekly report is drafted and verified number by number.
-
Step 5
Days 61–75: Narrow playbooks
Pre-approve two or three low-risk actions with limits, such as pausing ads with broken URLs; review every executed action weekly.
-
Step 6
Days 76–90: Measure and decide
Compare with the baseline and decide what to expand, keep or switch off, and which accounts or channels come next.
How do you measure whether the agent is worth it?
Measure it like a junior analyst: by the quality of what it flags, how fast it finds problems and the time it gives back, not by the number of recommendations.
| KPI | How to measure it |
|---|---|
| Alert precision | Share of alerts that led to an action or a documented decision |
| Time to detection | Hours from an issue starting to the alert |
| Recommendation acceptance | Approved ÷ total proposals, with reasons for rejections |
| Reverted changes | Approved changes undone within 30 days |
| Spend protected | Spend on broken pages, waste clusters or overpacing caught early, estimated conservatively |
| Hours saved | Weekly hours on checks and reports vs the baseline |
| Report accuracy | Number mismatches caught by verification or by readers |
| Guardrail incidents | Actions outside the playbook or limits; the target is zero |
Put the agent's running cost (model usage, hosting, maintenance hours) next to these numbers and review them monthly with the account owner.
Prompt library: five prompts we start from
These are task prompts, not the whole agent: they sit on top of system instructions, tools and guardrails, and every figure comes from a tool call. Adapt the thresholds to your accounts.
Daily triage:
Using only tool data for yesterday and the last 28 days, list at most five items that need attention today for {client}, ordered by money at risk. For each: what changed (metric, campaign, size), the most likely cause and the evidence you checked (CPC vs conversion rate, change history, search terms, landing page, CRM leads), the next step, and whether it needs approval. If nothing passes the thresholds, say so in one line. If a cause is unclear, say what you would check next.
Search term classification:
Classify each search term as: relevant, job seeker, DIY, free or cheap, wrong service, wrong location, competitor or informational. Return term, intent, confidence and a one-line reason. Never propose negatives for protect-list terms or terms with CRM leads. For non-relevant groups, propose n-gram negatives with match type and level, and list any active keywords each one would block.
Budget pacing and reallocation:
From the pacing and impression share tables, list campaigns projected to finish more than 10% over or under budget and explain each driver (change history, CPC, volume). Propose at most three moves, each capped at {max_change}% of the campaign budget, only into campaigns losing impression share to budget, inside target CPA and with no open tracking alerts. Give the expected effect as a range and flag campaigns in a learning phase.
Ad copy with brand rules:
Write 10 headlines and 4 descriptions for the {service} ad group in {city}, using the brand rules file, top converting search terms and customer objections provided. One idea per line, no forbidden claims, only active offers. Group lines by the hypothesis they test (price transparency, speed, trust). Do not count characters; the validator will return any line over the limit.
Weekly client report:
Draft the weekly update for {client} in their vocabulary ("booked jobs", not "conversions"): what happened, why, what we did, what we need from you. Use only figures from the report table and the approved changes log, citing the row for each. Under 250 words, plain English. End with the one decision we need from the client this week, if any.
Common mistakes in AI PPC management
- Automating before fixing measurement. If platform conversions do not match the CRM, every automated decision inherits the error.
- Letting the model do arithmetic. Pacing, CPA and projections belong in code.
- Too many alerts. Without materiality thresholds and deduplication, people stop reading.
- Changing budgets daily, which keeps Smart Bidding and Meta's delivery relearning.
- Shared logins for the agent, which kill the audit trail.
- Trusting Performance Max or Advantage+ reporting alone instead of CRM outcomes and controlled tests.
- Auto-sending client reports before the verification step has a long, clean track record.
The best AI PPC management setups are not the most autonomous ones. They are the ones where the agent does the watching, reading and drafting, and the people who know the business make the calls sooner and with better evidence. To see how we design AI agents for this kind of work, or to tell us how your paid media runs today, get in touch: we will tell you what we would automate first, and what we would leave alone.