Work

What if buying something properly did not mean operating five systems in the right order?

Procurement and business systems. Six years of it, and the closest of the three to being in front of users.

Nobody wakes up excited to create an RFQ. They want to buy properly, at the right price, without losing control of the process. The document is not the goal. The document is what the software made them produce in order to reach the goal.

The decision that matters is not how much to automate. It is where to put the guardrails. Automating a purchase is easy to demonstrate and hard to deploy, because the question a buyer is actually asking is what happens when it goes wrong, who sees it, and how quickly it can be undone. Guardrail placement is the design work. Automation percentage is a vanity number. I have worked out the ordering that question depends on in Trust Across Boundaries.

In 2021 I built a legal AI for contract management that identified and highlighted problematic clauses in procurement contracts. It combined NLP with an LLM for the language layer.

We called it a function. The word agent was not in use for this yet, and calling it one now would claim an autonomy it did not have and did not need. It read the contract, flagged the clauses, and a person decided.

The constraint was never speed and it was never the model. Penser met one edge of it. Vipani is still working the other: the distance between saying what you want and having it be done.

Approaching launch · building since 2020

What if saying what you want was enough to make it happen?

The gap this addresses is the one between wanting an outcome and performing the twelve steps that produce it. Stating the outcome is the natural act. Everything after it is the tax.

Products are in development across personal assistance, personal-brand management and specialised professional work, including a domain layer for law. Incorporated April 2026, on thinking two to three years older than that. I held off starting it because I wanted Vipani closer to launch first.

Building privately

What has to exist around intelligent models before we can trust them with more of the work?

The default answer to what happens to capability built inside an organisation is that it leaves with the person who built it. A consultant who tuned a workflow into something excellent takes the excellence when the engagement ends. The organisation paid for the learning and kept none of it. I think that is a solvable problem and the wrong default.

A deeper systems project, on the infrastructure questions that appear once AI systems need to remember, persist, coordinate and act over time. Incorporated May 2026, on thinking several years older, and well into engineering. Most of the work stays private for now. The coordination half of it is argued at length in Broken at the Seams.

Building privately

Penser Analytics, and a product called Sherlock

What if you could just ask your data a question?

A user types a business question in plain English. Natural language processing understands it and extracts client-specific modalities: products in retail, policies in insurance, campaign names in marketing. Those resolve into dimensions and measures, which are passed to a query engine that queries the client's own database. It runs the analysis and the machine learning models, gathers insights, and returns reports and insights.

Sherlock sat on top of an OLAP architecture where data was pre-summarised at several levels for fast analytical retrieval. Across the datasets we deployed it against, a request that had taken three to four weeks came back in under 53 seconds.

The natural-language layer is what people noticed. The retrieval architecture underneath it is what made the number possible.

The three to four weeks was not a slow team. It was how long an analytical request took to move through an organisation, from the person who wanted to know, to the person who could find out, and back again. Most of that time was queueing, clarifying and re-running, not thinking.

Insight generation was the kicker. Running a query quickly is a benchmark. Telling somebody which three of the results deserve their attention this week is the product, and it is the harder half.

Google Assistant was connected to Sherlock. A question spoken aloud produced an answer drawn from a company's own data, with no dashboard in between. Once the system underneath could answer, the interface stopped being the interesting problem. That the ease is allocated rather than removed is the argument of A General Theory of Human-System Interaction.

What it looked like in two rooms

A multi-channel retailer with more than a hundred stores and an online shop had a question that takes four seconds to ask and a long time to answer. What makes a customer buy online rather than in a store? Underneath it sit decisions about how each channel is positioned and how stores are run. They put the question to a large consultancy and to us at the same time.

Sherlock answered inside the meeting. Cohorts of customers who bought only online, only in store, and in both, set against what each cohort was worth, what each had cost to acquire, and what they actually bought. The pattern held: standardised goods and daily essentials went online, and anything a person wanted to handle first was bought in a store.

A few stores broke the pattern, selling almost nothing in person to the customers nearest them. The data could flag them; it could not say why. That took somebody driving out to look, and the answer turned out to be parking and the kind of neighbourhood each store sat in. The analysis found the anomaly in minutes and a person still had to stand in a car park to explain it.

The second was a personal care and cosmetics retailer with over a hundred and fifty stores across several formats, asking which campaign codes had worked. Their analytics team had the report ready before the meeting started. What happened in the meeting is that people began asking the second question, and then the third. Split it by demographic. Take out the discount hunters. Show me only the customers a coupon brought in.

Those questions got answered while they were being asked, which changed what the meeting was. Data stopped being the thing presented at the start and became part of the argument being had. That is a smaller claim than automation and a more useful one.

Autodata, and what COVID took

Autodata was a channel we sold through, and it was building into something substantial. Then COVID hit and it came apart, along with a good deal else that year.

Penser grew to fourteen people. It ended in 2020, when Vipani, the company I started next, acquired it.

The gap, and why the work moved

Natural language generation was the part we were not good at. Everything up to the finding worked. Saying the finding back to a person in ordinary language was the weak link in an otherwise working product.

Stanford was attacking NLG through systematic learning, grammar, phrases and rules. I believe Cambridge was too. Large language models took the other route. Trained on written text, they learned grammar by seeing it used rather than from a ruleset.

Two ways to learn chess. From the rulebook, or by watching a great many games until the patterns arrive. Even grandmasters study older players and their openings, so the two are not clean opposites.

The first widely available LLM arrived in June 2020. Three months after Penser closed and Vipani began. The next version of Sherlock my co-founder and I wanted is possible now, and analytics and data science systems which interact with an end user are still lacking.