AI agents & assistants
AI belongs in the parts of a process that involve reading and interpreting language — sorting messy inbound email, summarising a long document, drafting a reply. Everything else should be a deterministic workflow, because rules are cheaper, faster and more predictable.
The test we apply
Could a sufficiently detailed rule handle this? Then use a rule.
A model that is right 94% of the time is impressive and completely unacceptable for posting invoices. It is excellent for deciding which of six folders an email belongs in, where the cost of a mistake is that someone moves it.
Most of the "AI automation" being sold right now puts a model where an if statement belongs. It costs more per run, it fails unpredictably, and it is harder to debug.
Where an agent genuinely earns its place
- Inbound email triage. Classify by intent, extract the key details, set priority, route to the right queue, and flag anything urgent. The volume that never needed a human read is the win.
- Lead qualification from free text. A prospect writes three paragraphs about their situation. The agent extracts company size, budget signals, timeline and use case into structured CRM fields.
- Drafted replies for approval. Routine questions get a draft written against your own documentation, queued for a person to send with one click or edit in ten seconds.
- Long document review. Contracts, tenders and specifications read against your checklist, with the relevant clauses quoted and located rather than summarised away.
- Internal knowledge assistants. Retrieval over your own policies, SOPs and past projects, answering with citations to the source document so answers can be checked.
- Meeting and call summaries. Transcript into decisions, owners and dated action items, pushed into your task system.
Six rules we do not break
Humans approve consequences
Anything that spends money, signs something, or sends an external message keeps an approval step unless you explicitly decide otherwise in writing. Draft-and-approve is the default posture.
Structured output, always
Agents return typed, validated fields rather than prose that something downstream has to parse. If the output fails validation it goes to a review queue, not into your CRM.
Confidence thresholds
Every classification carries a confidence score. Below your threshold, it routes to a person. You choose where the line sits and we can move it once you see real numbers.
Evaluated before launch
We assemble a test set from your real historical examples and measure accuracy against it before anything goes live. You see the score. If it is not good enough, we say so.
Grounded in your documents
Knowledge assistants answer from your own content with citations, not from the model's general knowledge. If the answer is not in your documents, the correct response is "I do not know."
Logged and reviewable
Every agent decision is logged with its input, output and confidence, so you can audit what it did and we can improve it on evidence rather than anecdote.
What it costs and why
Agent builds start at $10,000 and run to around $35,000 for something spanning several systems with a substantial document corpus behind it.
They cost more than an equivalent workflow for a specific reason: the engineering is only half the job. The other half is assembling a real evaluation set from your historical data, measuring accuracy against it, and iterating on prompts and retrieval until the numbers hold. Skipping that is how you end up with a demo that impresses in a meeting and fails in week three.
Running costs are separate and usually modest — API usage for a triage agent handling a few hundred messages a day typically lands in the tens of dollars a month. We estimate it during scoping and you pay the provider directly.
Worked example
Before: a support inbox receives roughly 120 messages a day. Two people spend the first hour of every morning reading and sorting them into billing, technical, sales and spam, then assigning owners.
After: an agent classifies each message on arrival, extracts the account reference, sets priority from stated urgency and customer tier, and routes it. Messages it scores below the confidence threshold — typically the genuinely ambiguous ones — land in a single review queue.
The honest part: it will not get every one right. The measure that matters is whether the number of misroutes is lower than the current human error rate under time pressure, which we establish during evaluation before you commit to launching it.
Common questions
Do I need an AI agent or just a workflow?
If the task follows rules — when this happens, do that, move this data there — you need a workflow. It will be cheaper to build, cheaper to run, faster, and it will behave the same way every time. You need an agent when the task requires reading and interpreting language that arrives in an unpredictable shape. Most businesses need mostly workflows with a small amount of AI at the edges, and anyone telling you otherwise is selling something.
Which model do you use?
Usually Claude or GPT, chosen per task on measured accuracy and cost rather than preference. Classification and extraction often run well on a smaller, cheaper model; long-document reasoning needs a larger one. We build so the model can be swapped without rewriting the workflow around it, because this market moves quickly.
Where does our data go?
To the API provider you approve, under their commercial terms, which for both Anthropic and OpenAI business tiers means your data is not used to train their models. We can also work entirely with self-hosted open-weight models if your compliance position requires it, at the cost of some capability. This gets decided during scoping, not afterwards.
What if it gets something wrong?
It will, sometimes — which is why nothing consequential runs without approval, everything is logged, and low-confidence cases route to a person by design. The right comparison is not "perfect versus the agent" but "the agent versus a tired human doing the same task at 4pm on a Friday." We measure both before you launch.
Can you fix an AI agent someone else built?
Yes, and we are asked fairly often. The usual findings are no evaluation set, no confidence handling, no logging, and a model being used where a rule would do. We assess it, tell you whether it is repairable or better rebuilt, and quote either way.
Bring us the task, not the technology
Describe the work you think needs AI. We will tell you honestly whether it does, or whether a workflow at a third of the price would do the same job better.
Related: Workflow automation · Automation audit · Automation support