Menu Book a free audit
AI agents and assistants

Use AI where language arrives. Use a rule everywhere else.

An agent belongs where something has to be read and understood. Everything else should be a rule, because a rule is cheaper, faster and right every time.

Book a free audit 30 minutes, no chargeA written proposal in 2 business days
From $3,500 a chat agent 10 days or less Evaluated before launch
Support inboxRunning
Hi, our invoice 4471 is showing twice and the second one is the wrong amount. Can someone look before Friday?That is a billing duplicate. I have raised it with the accounts team with the invoice number attached, and flagged it for Friday.ClassifiedBillingConfidence0.94, above your line
AHandoverAnything under 0.80 stops and goes to a person with the thread
Models chosen per task, on measured accuracy and cost
  • Claude
  • GPT
  • Open weight
  • Self-hosted
  • Your documents
The test we apply

Could a detailed rule handle this? Then use a rule.

A model that is right 94% of the time is excellent for sorting email into six folders and completely unacceptable for posting invoices. Most of what is sold as AI automation puts a model where an if statement belongs.

What an agent is bad at

  • Arithmetic that has to be exact
  • Work where one wrong answer is expensive
  • Anything a written rule already covers
  • Running unattended with no confidence score
  • Being cheap per run at high volume

What an agent is good at

  • Reading free text that arrives in any shape
  • Sorting, extracting and summarizing
  • Drafting a reply for a person to approve
  • Finding the clause in a 60 page document
  • Answering at 9pm from your own material
Read

It classifies before anyone opens it

Intent, urgency, account and the details that matter, pulled out of a paragraph somebody typed in a hurry and written into structured fields.

  • Typed, validated output, never loose prose
  • Anything that fails validation goes to review
  • The account matched, not guessed
Extracted
IntentBilling questionUrgencyBefore FridayAccountMatched, 4471SentimentNeutral
Score

Every answer carries a number

Confidence is measured on every decision. Below the line you set it hands over rather than guessing, and you can move the line once you see real runs.

  • You choose the threshold, not us
  • Low confidence routes to a person with the thread
  • Input, output and score kept for every run
Confidence, last 500 runs
Handled without a person86%
Sent to a person14%Wrong after review0.6%
Approve

Nothing consequential sends itself

Anything that spends money, signs, or goes to a customer keeps an approval step unless you decide otherwise in writing. Draft and approve is the default, not an upgrade.

  • A draft queued for one click, not a surprise
  • Full thread and reasoning attached
  • Every action logged and reviewable
Waiting for you
Reply to Rachel MooreApproveRefund, $240Needs a humanBook the callAuto, allowed
Where an agent earns its place

Six jobs worth giving a model

Inbound email triage

Classified by intent, key details extracted, routed, and anything urgent flagged.

Lead qualification

Three paragraphs of free text become structured fields in your CRM.

Drafted replies

Written from your documentation, queued for one click to send.

Long document review

Contracts and tenders read against your checklist, with the clauses quoted.

Internal knowledge

Answers from your own policies and past projects, with citations to check.

Call summaries

Transcript into decisions, owners and dated actions in your task system.

Before you talk to anyone

Work out whether it is worth doing at all

An agent has to beat the person doing it now, not perfection. Start with what the task costs today.

That work currently costs you

$1,974

every month in wages alone, before the errors and the waiting.

A $3,500 build pays for itself in 1.8 months.

What it costs and why

A chat agent is $3,500, a voice agent $5,000, fixed before work starts. They cost more than a workflow because half the job is building an evaluation set from your own history and iterating until the accuracy holds.

See full pricing

Common questions

Do I need an agent or just a workflow?

If the task follows rules you need a workflow: cheaper, faster, the same every time. An agent is for language that arrives in an unpredictable shape. Most businesses need mostly workflows with a little AI at the edges.

Which model do you use?

Usually Claude or GPT, chosen per task on measured accuracy and cost. We build so the model can be swapped without rewriting the workflow around it.

Where does our data go?

To the provider you approve, on business terms that exclude training on your data. Self-hosted open weight models are possible if compliance requires it, at some cost in capability.

What if it gets something wrong?

It will, sometimes. That is why nothing consequential runs without approval, everything is logged, and low confidence routes to a person. The comparison that matters is the agent against a tired human at 4pm on a Friday.

Can you fix an agent someone else built?

Yes. The usual gaps are no evaluation set, no confidence handling, no logging, and a model doing a rule’s job. We tell you whether it is repairable or better rebuilt.

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.

You will talk to Fayez or Kareem. Nobody else.

Related: Workflow automation · Scoping and proposal · Automation support