ContactBook a call

Systems

Six months of AI agents in production: what broke

The model held up. A renamed field, a drifting confidence score, a new invoice format and an unread Slack channel did not. What failed and what we changed.

Cover for “Six months of AI agents in production: what broke”, with the line: Every failure we logged was in the plumbing, the data or the people. None was the model being wrong.
Every failure we logged was in the plumbing, the data or the people. None was the model being wrong. Image: SBD.

When clients ask what goes wrong with agents, they expect a story about the model hallucinating. We have been running agent steps inside client systems for lead qualification, invoice intake, support triage and order status since the spring, and the model is almost never the thing that broke.

Here is what did.

The short answer

In six months the failures that mattered were a renamed CRM field, a supplier changing its invoice layout, a confidence threshold that drifted as the prompt was tuned, and alerts going to a channel nobody read. All four were caught late because the agent kept reporting success. The fixes were monitoring outcomes, validating data shape at the door, re-testing the threshold on every prompt change, and naming an owner. The model stayed the same throughout.

Four failures, none of them the model: from SBD client systems, April to October 2026
Four failures, none of them the model. From SBD client systems, April to October 2026.

What we run

Across clients the agent steps look like this: an agent reads something unstructured, decides, and writes one structured record. Workflows on n8n or in code do everything around it. The lead enrichment system is the published example. The others are in manufacturing, distribution and services and are not yet written up.

So what follows is what we saw, rather than a benchmark.

Failure one: the field that changed its name

This was the most common failure by a distance. A client’s CRM admin renamed “Email” to “Work email” for tidiness. The agent’s output mapping still pointed at “Email”. The workflow wrote the record, the write succeeded, and the email field was blank for nine days. Follow-ups went nowhere.

What we changed. Every workflow now validates the shape of the record at entry and exit. A missing required field stops the run and lands in a review queue with the raw payload. We also subscribe to the CRM’s schema change events where the platform offers them.

Failure two: the supplier with a new template

The invoice agent had an extraction accuracy we were happy with. Then one supplier, who sent about a third of the invoices, moved to a new accounting package. Totals moved to the top of the page, line items to a second page. The agent kept extracting, with the wrong total, and the arithmetic check did not catch it because the line items were missing too.

What we changed. The validation now requires line items to be present and to sum to the total, and a mismatch or an absence routes to a person. We also log per-supplier confidence, so one supplier’s numbers dropping stands out against the rest.

Failure three: the threshold that drifted

We tune prompts, and each round of tuning changed how the model expressed confidence. A threshold set at 0.8 in April meant something different by July, and by then the agent was auto-approving leads it should have escalated.

What we changed. A prompt change now triggers a replay of the last two hundred decisions against the new prompt, and the threshold is re-set from that replay before the change goes live. The drift came from how we used the model.

Failure four: the channel nobody read

Every alert went to a Slack channel called #automation-alerts. Three people were in it, and all three had muted it by June. The failure above sat in that channel for a week.

What we changed. Each workflow has one named owner and a backup, and alerts go to the owner directly. The owner’s name is in the handover document and in the alert text.

What did not break

  • The model. We changed providers once for cost and accuracy did not move.
  • The orchestration layer, once retries and queues were in place.
  • The review queue. Every client’s team used it, and it is the part they mention when asked if they trust the system.

The pattern

Every failure was a silent success. The agent did its job on bad input, or did its job and reported to nobody. That is why most of our effort now goes into monitoring.

If you are running agents and have not had one of these yet, there may be one you have not found. The Audit looks for these four, and the fixes above are standard in every build.

Questions we get asked

Which model do you use?

Claude or OpenAI through their business APIs, chosen per task on cost and format reliability. In six months the choice has mattered far less than the validation around it.

How often do agents escalate to a person?

It depends on the task and how long the prompt has been tuned. A well-tuned qualification agent escalates a small minority of leads. Invoice intake escalates more, because the input varies more. The number should fall over time, and if it rises, something upstream has changed.

Could a client run this without you?

Yes, and some do after hypercare. The handover document lists every workflow, its owner, its validation rules and its alerts. That document is what makes it runnable.

Sources and further reading

  • AI agents
  • Operations
  • Lessons

Keep reading

Cover for “When to replace a pile of Zaps with an agent”, with the line: From single-track automations to agents that hold state and route exceptions.

Systems

How to move a Zap-based process to an AI agent

A step-by-step guide to replacing a chain of Zaps with an agent: inventory, data model, approvals, side-by-side running, switch-off and pitfalls.

Cover for “Poppy 1.0 and what personal agents will ask of you”, with the line: Publish the basics, then add actions with approval.

Systems

Poppy 1.0 and what personal agents will ask of you

Sierra released Personal Agent Protocol (Poppy) 1.0 on 13 Nov 2026. What a mid-size business off Shopify should expose to personal agents, and in what order.

Cover for “Moving off OpenAI Agent Builder before 30 November”, with the line: Agent Builder and Evals shut on 30 November 2026. ChatKit stays. Export what you need before then.

Systems

Moving off OpenAI Agent Builder before 30 November

OpenAI Agent Builder and Evals are due to close on 30 November 2026. What to export now, where agents can go, how to move them and which ones to retire.

Find out what is worth automating

About 95% of AI pilots never reach the profit line. We start with the work your team still does by hand and measure each build by the hours it saves.

See how we work

Cookie preferences

Choose what this site may use on your device. You can change this at any time from “Cookie settings” in the footer.

Strictly necessary

Keep the site working and remember the choices you make here. These cannot be switched off, and none of them track you.

Always on
What this includes (4)
NameProvider PurposeDuration
sbd-consentThis site Remembers your cookie choices12 months
Vercel Web AnalyticsVercel Inc. Counts page views in aggregate. Sets no cookie and stores nothing on your deviceNone
ticker-dismissed, sbd-note-offThis site Remembers that you closed a notice or language tipUntil the tab closes
sbd-voice-resume, sbd-voice-prefillThis site Carries a voice conversation across pages, only if you start oneUntil the tab closes

Analytics

Tell us which pages are read and where visitors leave, so we can improve them. IP addresses are anonymised and nothing is used for advertising.

What this includes (2)
NameProvider PurposeDuration
_gaGoogle Analytics (Google LLC) Distinguishes visitors with a random ID13 months
_ga_FPXDMQPRV1Google Analytics (Google LLC) Keeps the session state for this site13 months

Functional

Embedded services that set their own cookies. Without them the site still works; you get a link to open the service instead.

What this includes (1)
NameProvider PurposeDuration
Booking calendarCal.com Inc. Shows live availability and books the call on the booking pageSet by Cal.com

Advertising

Not used. This site has no advertising, retargeting or social media tracking, and we never sell data.

Not used

Questions about your data: hello@smallbaydigital.com