AI that gets used
Most AI projects fail quietly. Here is how to be one of the ones that does not.
The pattern in a 50–200 person firm is depressingly consistent: a strategy deck, a pilot nobody logs into, a licence renewal nobody can justify, and a quiet decision never to mention AI again.
I am not selling you a model. Models are a commodity and getting cheaper. What I sell is knowing which part of your business will actually absorb one — which is a question about how your business runs, not about technology.
Why they fail — and what I do differently
Four reasons account for nearly all of it. None of them is the model being bad.
It solved a problem nobody had
Somebody picked the use case from a conference talk rather than from the work. The tool is impressive and irrelevant.
What I do: start from where your people are losing hours or making mistakes. If I cannot point at the specific task, I will not build the thing.
The data was not ready
The information lives in four places, in three formats, half of it in somebody's email. AI over a mess produces confident nonsense.
What I do: get the information in order first. It is unglamorous and it is why the builds work — it is also the same discipline as the document control work on the rest of this site.
Nobody changed how they worked
The tool sat beside the job instead of inside it. People went back to the old way within a fortnight because the old way was still there.
What I do: put it where the work already happens, and train the people who have to live with it. Adoption is the deliverable, not the software.
It could not reach the real systems
The useful data sits in a system built in 2011 with no API and a vendor who has disappeared.
What I do: this is the part I am unusually good at. Four technology eras means I can generally get a modern tool talking to an old one without a replacement programme — and I will say honestly when the old thing genuinely has to go first.
Where AI genuinely pays
Judged on what has actually worked in real businesses, not on what demonstrates well.
- Reading more than a person can read. Contracts, tenders, submittals, policies, correspondence — anywhere the volume means a human samples rather than reads
- Catching what a tired person misses at 6pm. The same check applied the same way every time, without fatigue or an approaching deadline
- Answering questions from your own material. Twenty years of documents becoming something you can interrogate rather than search
- First drafts. Proposals, reports, correspondence in your own format and tone — a good starting point beats a blank page
- Structured research. Markets, counterparties, competitors, prospects — gathered and summarised to a consistent shape
- Turning conversation into action. Meetings into decisions, owners and dates, without anybody writing minutes
Notice what is common to all six: the work still gets checked by somebody who knows the business. That is not a limitation to design around. It is the design.
The systems I have built
Working systems, in use, not concepts. I will run whichever is closest to your problem live on a call — that is the fastest way for you to judge whether I build things that actually work.
For the business as a whole
- Operating systems built around a whole business — domain agents under one orchestrator covering programme, risk, cost and compliance, working from one set of data rather than four versions of the truth. Built for a contracting group where the alternative was another decade of spreadsheets
- Private knowledge assistants — your own documents, searchable by question rather than by filename, with the source attached to every answer
For winning and doing the work
- Tender and bid intelligence — reads a bid pack the way an experienced estimator would: scope gaps, pricing conflicts, onerous terms, and a benchmark against what the job should cost. On a single tender it identified HK$1.5m of potential costs to be avoided
- Document and contract review — bulk reading with only the exceptions surfaced to a human
- DealMind — a negotiation coach that prepares you for the specific deal in front of you, maps what each side actually wants, and tells you where your position is weaker than you think
For growth
- Prospect research and qualification — finding and sorting the people genuinely worth approaching, rather than a bigger list
- Fundraising support — investor targeting, materials, and the follow-up discipline that is usually where it falls down
- Proposal and content systems — first drafts in your own voice and format
For getting your firm properly onto AI at all
- Adoption programmes — Claude, Copilot or Gemini set up properly: connectors into your real systems, guardrails on what data goes where, role-based training, and someone to ask afterwards
- Guardrails and policy — written in English, so your people know what is allowed before somebody pastes a client contract into a chatbot
Websites and app development
The other half of this. AI has not only given us assistants — it has changed what it costs to build a small piece of working software.
Things that were a six-figure development job three years ago are now days of work. That is genuinely new, and it means the tool you priced up and shelved is worth looking at again.
What I build
- Client portals and customer-facing tools that make a small firm look substantial
- Quote, cost and specification calculators
- Internal trackers and approval flows that replace shared spreadsheets
- FileMaker and custom database applications — new builds, and rescues of ones that have outgrown themselves
- Integrations between systems that were never designed to talk
- Business websites, from a single landing page upward — thirty-plus built
How I keep it honest
Fast building is easy to do badly. Software that works on the demo and falls over in month three is worse than no software, because now it is load-bearing.
I have been writing and buying software since long before this was easy, which mostly means I know which corners not to cut — and where an off-the-shelf product would serve you better than anything custom, I will say so and lose the build.
Scoped, priced and agreed in writing before anything starts, like everything else here.
Built by Innov8
DealMind
An AI negotiation coach — preparation before the room, and support during it. In use now.
Most negotiation training is a two-day course you have forgotten by the following month. DealMind works the other way round: it prepares you for the specific deal in front of you, maps what each side actually wants, and tells you where your position is weaker than you think.
It is also a fair example of what I mean by all of the above — a bounded tool, built to be used in a real situation, rather than a platform looking for a purpose.
Where it does not work yet
Worth saying plainly, because the market will not.
AI is poor at anything where being roughly right is worse than being no help at all. It will not sign off a structural calculation. It will not make a judgement call about a person. It will not replace somebody who knows your business and your clients.
It is bad at knowing what it does not know, which is why every system I build shows you its source. It is bad at edge cases in your particular business, because your edge cases are not in its training data — they are in your people's heads, and getting them out is part of the job.
And it will not fix a process that is broken. Automating a bad workflow gives you a faster bad workflow.
On headcount. If what you want is to cut people, I am probably the wrong call — the savings usually turn out to be smaller and slower than promised. If what you want is your existing people covering more ground with fewer mistakes, and getting to the work only they can do, that I can do and have done.
Your data
Plain answer: your data is yours, it is not used to train anybody's model, and I will tell you before anything leaves your environment.
The full position is in the AI Policy — written in English rather than legal, and governed by the Hong Kong Personal Data (Privacy) Ordinance.
Want to see one of them running?
Book a quick call and I will demo whichever system is closest to the problem you have.