AI, honestly
The AI question, answered in two weeks not two quarters
We find where AI pays in your business, where a plain rule would do the job better, and what has to be fixed first. It comes out of the same diagnostic, at the same fixed fee.
The usual offer
What a six month AI engagement usually produces
A discovery phase, a maturity assessment, a workshop series, and a roadmap. The roadmap is often sensible. The problem is that by the time it lands, it has cost more than the first year of benefit, and it recommends work nobody has budget left to do.
It also tends to assume the underlying processes are sound, because nobody looked at them. They usually are not, and that is where the real constraint sits.
We would rather spend two weeks showing you the three things worth doing, with a price against each, and let you decide.
What you get from us instead
- A ranked list of where repetition actually costs you, in hours and loaded cost
- For each one, whether it needs a model, a rule, or a fixed handover
- What has to be cleaned up before any of it is worth switching on
- A price against each item, so you can take one, take all, or take none
- An explicit list of what we think you should not automate
How we look at it
Three questions, asked in order
Most AI projects fail on the second and third question, having never asked the first.
Where is the repetition
Which steps a person repeats every week, how many minutes each one takes, and what that costs over a year. Measured, not estimated in a workshop.
What is blocking it
Usually missing data, an approval that was never actually defined, or two systems that disagree about the same record. None of that is solved by a model.
What is worth doing
Some of it is worth automating. Some of it is worth deleting. Some of it should stay manual because the volume does not justify the risk.
Rule or model
Half of what gets sold as AI is not AI
This is the table we end up drawing on a whiteboard in almost every first conversation, so here it is in advance.
| What you want | What it actually needs | Why |
|---|---|---|
| Stop invoices going out at the wrong rate | A rule | Effective-dated rates and one source of truth. Deterministic, auditable, and it cannot be wrong. |
| Chase missing timesheets | A rule | A scheduled job and a reminder. There is no judgement involved, so there is nothing for a model to add. |
| Answer "what is in the pipeline for Q3" | A model | The question is asked a hundred different ways. Language is the hard part, and the data is already structured. |
| Read a client contract and pull out the rate card | A model | Unstructured input, high variation, and a human checks the output before it is used. |
| Summarise account history before a renewal call | A model | Synthesis across notes, cases and emails. Genuinely faster than a person, and low cost if it is wrong. |
| Decide which candidate to submit | Neither, yet | The decision carries real consequences and your historical data encodes past bias. Assist the person, do not replace them. |
| Reconcile two systems that disagree | Fix the join | Automating a reconciliation is paying forever to paper over a defect that could be fixed once. |
Shipped, not slideware
Three things we have actually put into production
Described without client names, because the work was done under agreement.
Pipeline questions answered in chat
An assistant that answers questions about live CRM data in plain language, so a manager stops waiting on a report request for something they needed in the next ten minutes.
What changed: routine report requests to the operations team dropped sharply, and the questions that remained were the ones actually worth a person's time.
Account history before a renewal call
Relevant history surfaced automatically ahead of the conversation, rather than someone scrolling three record types for twenty minutes beforehand.
What changed: preparation time per call fell from roughly twenty minutes to under five, and nobody walked in unaware of an open complaint.
AI in the development pipeline
AI drafts the design, the code and the tests. A senior engineer remains a mandatory gate on every change, and nothing reaches production without that review.
What changed: the drafting stage got materially faster while review stayed human, which is the part most teams get the wrong way round.
The honest part
What we will tell you not to do
The diagnostic includes an explicit list of things we think you should leave alone. It is usually the most useful page in the document, and it is the part a vendor with a licence to sell cannot write.
We do not resell any platform and we take no commission from one, so there is nothing in it for us either way.
See a sample findings pageThings we have advised against
- Automating a process that was about to be retired anyway
- A chatbot over a knowledge base nobody had maintained in two years
- Predictive scoring on eleven months of inconsistent history
- Replacing a judgement call that happens four times a month
- Any model sitting on top of two systems that still disagree
Find out in two weeks, not two quarters
A fixed fee from USD 3,500, a ranked list you own outright, and an honest answer about what is not worth doing.