Checkpoint 8 · Automation, applied to unstructured work
Language models applied to the unstructured work — intake, triage, summarising, drafting, extraction — with a human reviewing anything a customer sees and a route back to a person when it fails.
Deterministic automation handles work with a known shape: this field moves to that system, this trigger sends that reminder. A great deal of the work in a service business does not have a known shape. An enquiry arrives as three paragraphs of prose. A call produces twenty minutes of speech. A supplier sends a PDF laid out differently every quarter. That is the work language models are genuinely good at, and it is where this service applies them.
It applies them narrowly and on purpose. A model is a component with a failure mode — it produces a fluent answer whether or not it is a correct one — so the design question is never just "can it do this" but "what happens the day it is confidently wrong, and who notices". Anything a client sees passes a person first unless the cost of being wrong is genuinely trivial.
Ariadne builds these the way the rest of the system is built: the process is designed before anything is automated, every path reports success and failure, and a human route stays open. Where a deterministic script would be cheaper and more testable, that is what gets built instead, and the reasoning is written down.
Fit
Symptoms
Scope
Inbound free text read, categorised against your criteria, given a priority, and routed to the right person with the context attached. The classification is recorded next to the enquiry, so when it is wrong you can see that it was wrong and why.
A first reply prepared from the enquiry and your own material, waiting in a queue for someone to approve, edit, or discard. The gain is the blank page, not the send button. Approval rates are tracked, because a draft that always gets rewritten is not saving anything.
Structured fields pulled out of documents that arrive in an inconsistent layout, with a confidence signal and a review step for anything below the threshold. Extraction is checked against a set of your real documents before it goes live, not after.
Summaries and next actions written back to the right record in the CRM, so context survives the conversation. Where recording touches confidential or health information, consent and retention are settled as part of the design rather than assumed.
Retrieval over your documented policies, prices, and past answers so replies are grounded in your material rather than in a model’s general impression of your industry. Answers cite the source document, which is what makes a wrong one findable.
A test set of real examples with known-good outcomes, run before launch and again whenever the prompt or the model changes. Every call logged with its input, output, and cost. Alerting on failure and on spend. A documented fallback to a person on every path.
Process
Frequency, time cost, error cost, and whether the output can tolerate being probabilistic. Tasks that fail that last test are sent back to deterministic automation or left alone.
Twenty to fifty real examples from your business with the outcome you would have wanted. Without this there is no way to tell an improvement from a change, and no way to notice a regression.
What a person sees, when they see it, and what happens if they do nothing. The default is that customer-facing output is reviewed and unreviewed output stays internal.
Logging, failure alerts, a spend ceiling, and a manual route that is tested rather than assumed. Then the evaluation runs again on live examples to see whether it held.
Outcomes
These are the kinds of change this work produces. No numbers appear here, because no client result has been measured and published yet. When one has, it will appear in Work with its baseline, method, and time window.
Tooling
Categories, not endorsements. Model choice follows the task, the data boundary, and cost — and where a deterministic rule is cheaper and more testable, it wins.
The connected view
A growth system is only as strong as its weakest handoff. Fixing one checkpoint in isolation helps only if the adjacent checkpoints can carry the result.
See the whole systemCheckpoint 9 feeds evidence back into checkpoint 1. That loop is the point.
Where this sits in the system
Before this: which tasks are worth doing, what each costs to run, and where the confidentiality boundary sits.
AI auditAlongside this: the deterministic half of the same job, which is usually the larger and cheaper half.
Business automationAfter this: classified, routed enquiries are only useful if the pipeline behind them has an owner and a next action.
CRM and sales systemsStraight answers
If yours is more specific, put it in the audit form.
support@ariadne.fyiBusiness automation handles work with a known shape — a field moves, a reminder fires, two systems stay in sync — and it is deterministic, testable, and cheaper to run. This service covers work with no fixed shape: prose, speech, and documents that arrive differently every time. Most projects need both, and the split between them is a design decision made per task rather than a preference.
It will, so the design assumes it. Customer-facing output is reviewed by a person before it goes out. Internal output carries a confidence signal and a review threshold. Every call is logged with its input and output so a wrong answer can be found and traced rather than argued about. And every automated path keeps a route back to a human that has actually been tested.
Partly, and the honest answer is that it depends on the task and the tolerance. What can be decided is which information may leave your systems, what is redacted before it does, which vendor terms are acceptable, and what runs on infrastructure you control instead. Those are decisions that belong to you, written down before anything is built. Where a professional obligation or privacy law constrains the answer, that constraint sets the boundary rather than the technology.
Whichever fits the task, the data boundary, and the budget, and it is written down so it can be changed. Models are replaced far more often than the systems around them, so the evaluation set, the review points, and the logging are the durable parts. Being able to swap a model and re-run the evaluation is worth more than any current benchmark ranking.
There is a usage cost per call and a labour cost for review, and both belong in the business case before the build. Usage is metered, capped, and alerted on, so a change in volume shows up as an alert rather than as a surprise invoice. Where review time would cost more than the work it replaces, the correct recommendation is not to build it.
These first-party sources support the platform, policy, and regulatory statements on this page. Ariadne’s recommendations and planning heuristics are identified separately.
Your practical starting point
The audit ranks the repeated work by what it costs you and says which parts a model should handle, which parts a script should handle, and which parts should stay with a person.