Build an AI Reporting Agent, or Buy a Weekly Read?

Build or buy marketing reporting: 14 of 87 agencies built their own and kept it. Here is what a DIY Meta Ads agent costs to maintain, and when buying wins.

By Alex Neiman·Nov 6, 2026·10 min read

Build or buy marketing reporting is a maintenance question, not a build question. Getting an AI agent to produce its first useful Meta Ads read takes a weekend. Keeping it correct through token expiry, API version sunsets and numbers that get restated for 28 days is the part that never ends.

That second part is where the decision actually gets made.

Why this matters right now

Vendors have started publishing build-your-own tutorials. Polar Analytics shipped a guide to building a daily Meta Ads review agent. Supermetrics has been building out an agentic workflow lane. Whatagraph went and surveyed 87 agencies on the question directly.

The tutorials are honest about the build. They are quiet about the upkeep. And the upkeep is the whole cost.

What 87 agencies actually reported

Whatagraph surveyed 87 marketing agencies between September 2025 and October 2026 on whether they built or bought their reporting stack. The headline is not "building is a trap." It is narrower and more useful than that.

FindingNumberWhat it means
Built their own and were satisfied14 of 87Roughly one in six. Building works for a real minority — not nobody
Pursuing a hybrid model11 of 87Buy the data layer, build the workflow on top
Raised AI accuracy problems14 of 87Wrong numbers from an agent with no governed data layer underneath it
Initial build, one estimate400 hoursA Dutch agency's own figure for getting there internally
Initial build, contractor route$10,000–$12,000A US agency's spend, which got them "65 to 70% of the way there"
Ongoing maintenance5–30 hrs/monthOne head of strategy, on Looker Studio alone
Metrics deprecated in 2025140+Across major ad platforms. Every one is a possible break

Two numbers do the real work here. $10,000 to $12,000 bought 65 to 70% of a stack — the remaining third is the hard third. And five to 30 hours a month is the range for maintaining one dashboard layer, not an agent, not a pipeline, not a warehouse.

The five costs nobody budgets

Every line below is a documented behaviour of the Meta APIs, not a hypothetical. These are the things a DIY agent has to survive, and they are the reason the maintenance number is a range rather than a point.

#CostThe documented behaviourWhat it breaks
1Token refreshLong-lived tokens "last about 60 days," and Meta explicitly warns not to depend on thatSilent auth failure roughly every two months, per account
2Version expiryMarketing API v24.0 released Oct 8, 2025 and expired Oct 6, 2026. v23.0 expired Jun 9, 2026. Current is v25.0A working agent stops working on a date you did not write down
3Rate limitsFull Access caps at a score of 9,000; reads cost 1 point, writes cost 3; decay is 300 secondsPartial pulls that look like real performance drops
4Keeping your tierFull Access needs 500+ calls in the past 15 days and an error rate under 15% over your last 500 callsA quiet month or a buggy deploy can cost you the tier itself
5Restated numbers"Insights refresh every 15 minutes and do not change after 28 days of being reported"Yesterday's number is provisional. Your agent already acted on it

Cost 5 is the subtle one, and it is specifically a problem for daily agents. Meta's own best-practices guidance is explicit:

"Insights refresh every 15 minutes and do not change after 28 days of being reported." — Meta for Developers, Marketing API Insights Best Practices

Read that as a warranty with a 28-day term. An agent that compares yesterday against a trailing seven-day average is comparing one unsettled number against seven other unsettled numbers. Both sides of the comparison are still moving. The agent will not tell you that, because it has no way to know.

Cost 4 is the least obvious. The access tier is not granted once — it is scored continuously against your last 500 calls, so a deploy that throws errors for an afternoon is not just a bad afternoon. The full rate-limit arithmetic is in Meta Marketing API rate limits.

The tell in the build tutorials

Read Polar Analytics' build guide closely and the interesting part is the prerequisite list, not the instructions.

To build their daily Meta Ads review agent you need: a data source that already holds Meta Ads and Shopify data together with MER, blended ROAS and CAC already defined, a paid Claude plan with the Polar MCP connector enabled, a scheduler, and Slack or Notion to receive the output.

So the "build" is a prompt and a cron entry sitting on top of a purchased data layer with pre-governed metrics. That is not building a reporting stack. That is buying the hard part and building the easy part — which is precisely the hybrid model that 11 of the 87 agencies described.

The guide is also candid about what never ends. It tells you to treat each rejected recommendation as "a rule to add or tighten," to run a deeper weekly review, and to keep refining prompts against your approval patterns. That is a standing weekly commitment, written into the instructions.

This is a different question from the three adjacent ones, and worth keeping separate:

This post is narrower: you, building it yourself, versus paying someone for the finished read.

Running the numbers

The arithmetic below applies a $120/hour blended rate to the figures the 87-agency research reported. These are not observed outcomes from any one account — they are that survey's numbers, costed out, so the crossover point is visible.

Line itemBuild (floor)Build (midpoint)Buy
Initial$11,000 contractor400 hrs internal ≈ $48,000$0
Completeness after initial65–70%Assumed complete100%
Monthly upkeep5 hrs ≈ $60015 hrs ≈ $1,800$500 per ad account
Initial amortized over 24 mo≈ $458≈ $2,000$0
Effective monthly, year one≈ $1,058≈ $3,800$500 × accounts

The crossover lands between roughly 2 and 8 ad accounts, depending entirely on which maintenance number you believe. That is a narrower buy window than most vendor comparisons admit, and it is worth stating plainly: on a 15-account roster, a competently built stack is cheaper on cash than per-account pricing. Below the crossover, buying is cheaper on cash and much cheaper on attention. Above it, a build pays for itself on cash — assuming the upkeep hours stay at the floor, which is the assumption the research says breaks most often.

Note what the floor column does not include: the missing 30%. The agency that spent $12,000 got to 65–70%, and the last third is the governed metric definitions, the retry logic and the version pinning — the three things that cause the silent failures above.

When building actually wins

Fourteen of 87 built their own and kept it. They were not wrong. Here is the honest read on who they are.

Build when:

  • You have an in-house data engineer whose job already includes pipeline upkeep. Not a willing marketer. An engineer with this in their remit.
  • Your metric definitions are genuinely proprietary — a blended efficiency target or a contribution-margin rule no vendor models.
  • You are on more than ~8 ad accounts, where fixed build cost amortizes against per-account pricing.
  • You need the warehouse anyway for non-advertising data, so the connector work is already funded.

Go hybrid when:

  • You want custom workflows but not custom plumbing. Buy the governed data layer, build the routing and the rules on top. This is what the Polar tutorial actually describes, and what 11 of 87 agencies chose.

Buy when:

  • Nobody on the team owns API maintenance as a named responsibility.
  • You are on one or two accounts, where the amortized build never catches up.
  • What you need is a decision, not a dataset. This is the real split — see Meta Ads reporting software for the per-account version, or reporting for agencies for the roster view.

That last point is the one that gets skipped. A DIY agent produces a document. Somebody still has to read it, decide which of its suggestions to trust, and act. Good Morning's output is an urgency-tiered action list — Act today, This week, Monitor — already diagnosed before it reaches you. Zero analysis required at the other end. The build-vs-buy comparison is usually framed as data pipeline versus data pipeline, when the thing being compared is really a dataset you still have to interpret versus a set of action items you just execute.

Common mistakes

  1. Costing the build and not the calendar. The initial number is the one people estimate. The 5–30 hours a month is the one that decides the outcome.
  2. Calling a vendor tutorial a build. If step one is "connect your purchased data source with metrics already defined," you are pricing a hybrid, not a build.
  3. Trusting a daily agent's day-over-day deltas. Insights do not settle for 28 days. Daily comparisons on unsettled data generate confident noise.
  4. Not pinning an API version — or pinning one and not tracking its expiry. Marketing API v24.0 expired on October 6, 2026. Every version has a date like that.
  5. Putting an LLM on ungoverned data. 14 of the 87 agencies reported accuracy problems from exactly this. The agent will state a wrong number in the same tone as a right one.

FAQ

Is it cheaper to build or buy marketing reporting? On one or two ad accounts, buying wins on cash and on attention. Above roughly eight, a build amortizes on cash alone. In between, it depends on whether your maintenance lands at five hours a month or 30 — and the 87-agency research suggests most teams guess low.

How long does a DIY Meta Ads reporting agent take to build? One agency in the research estimated 400 hours internally. Another spent $10,000–$12,000 on contractors and reached 65–70% completion. A basic daily agent on top of an existing governed data source is much faster — that is the hybrid path, not a build.

What breaks most often in a self-built Meta Ads pipeline? Authentication and API versions. Long-lived tokens last about 60 days and Meta warns against depending on that. Marketing API versions expire on fixed dates — v24.0 expired October 6, 2026. Both fail silently.

Can I just point Claude or ChatGPT at my ad account and skip all this? You can, and it is worth testing. But an LLM over ungoverned data is where the accuracy complaints in the research came from. The setup specifics are in Meta Ads MCP reporting with Claude.

Does buying mean I lose customization? Partly, and that is the real trade. The hybrid model exists for this reason: buy the data layer and the diagnosis, build the routing and the rules you genuinely need. Comparisons against specific vendors are at Polar Analytics, Whatagraph and Supermetrics.

The short version

Token expiry, version sunsets, tier scoring and 28 days of restatement are the platform's documented default behaviour, and they arrive on their own schedule whether or not anyone is free that week.

If nobody on your team owns that calendar, buy the read. Meta Ads reporting software that hands you a pre-diagnosed action list costs $500/mo per ad account and has no maintenance window. If you do have an engineer and more than eight accounts, build it — and budget the upkeep at 30 hours a month, not five.

Sources

  1. Whatagraph — Should you build or buy reporting stack? Insights from 87 agencies (87 agencies surveyed September 2025–October 2026; 14 built their own and were satisfied; 11 pursuing a hybrid model; one agency estimated 400 hours to build; another spent "$10,000 to $12,000" on contractor work and got "65 to 70% of the way there"; one head of strategy spends "between five and 30 hours a month just maintaining Looker Studio"; 140+ individual metrics deprecated across major ad platforms in 2025; 14 of 87 raised AI accuracy problems without a governed data layer)
  2. Polar Analytics — How to Build an AI Agent That Reviews Your Meta Ads Every Day (requires a data source holding Meta Ads and Shopify data with MER, blended ROAS and CAC already defined, a paid Claude plan with the Polar MCP connector, scheduled tasks, and Slack or Notion for output; "Each rejection is a rule to add or tighten"; "Your team reviews the list and approves any budget change")
  3. Meta for Developers — Access Tokens (short-lived tokens "typically last about one to two hours"; long-lived tokens "last about 60 days"; "Do not depend on these lifetimes remaining the same — they may change without warning or expire early")
  4. Meta for Developers — Graph API and Marketing API Versions (Graph API v26.0 released July 29, 2026; Marketing API v25.0 released February 18, 2026; Marketing API v24.0 released October 8, 2025 and expired October 6, 2026; v23.0 expired June 9, 2026)
  5. Meta for Developers — Marketing API Rate Limiting (Development/Limited Access max score 60, Full Access max score 9,000; "A read API call is equal to 1 point, and a write API call is equal to 3 points"; "The decay rate is 300 seconds"; X-Ad-Account-Usage and X-FB-Ads-Insights-Throttle headers)
  6. Meta for Developers blog — Update to Ads Management Standard Access (effective May 4, 2026; Advanced Access renamed Full Access; "500+ Marketing API calls in the past 15 days"; "Error rate < 15% in the last 500 calls" calculated over a rolling window rather than a fixed time period)
  7. Meta for Developers — Marketing API Insights Best Practices ("Insights refresh every 15 minutes and do not change after 28 days of being reported"; "Insights metrics may continue to update for a couple of days after an ad has completed")

Related reading