Garry Jakubiak
Menu

My job · IoD Solutions · since January 2025

Taking AI from the demo to real work

At IoD Solutions, I build the Industry practice: the AI consulting offer for industry.

My role combines building the offer with delivering it for clients, through to adoption by the teams.

A few projects

On these projects I handled sales, scoping and project management; the engineering is the team’s.

Industrial quotes prepared without re-keying

CGR International · in production

The problem
Preparing a quote meant reading very different technical drawings, then re-keying dimensions, loads, tolerances and materials into the calculator.
What we did
An application that reads the drawings with a vision model and feeds the calculator directly; each material written as free text is matched against a catalogue of 288 grades; the data stays in the European Union.
What came of it
Extraction accuracy measured field by field: 85 to 96% depending on the documents.
What I take from it
Measured with our evaluation tool: Gemini’s “Flash” model does better than the most powerful one, for far less. The biggest model is not always the right one.

Four agents to respond to calls for tender

Confidential client · in testing

The problem
Responding to a call for tender takes days, and inconsistent data can get the bid rejected.
What we did
Map the process with the team, build four lightweight agents in Copilot, the tool they already use, test them, then share them with the team.
Results so far
Roughly one working day saved per tender, and more consistent data. This is an estimate from the testing phase, not yet a measurement in production.
What I take from it
Build the agents with the people on the ground from day one, to encourage adoption: bottom-up innovation.

Also: reports supporting research tax credit claims (France’s Crédit d’Impôt Recherche), drafted from a client’s own documents (Vermon).

Four ways to step in

Every company starts from a different place. We pick the level of intervention that fits the problem.

  1. Raise awareness and trainEnable teams to work independently on their own use cases, and to judge what the AI produces.
  2. DiagnoseStart from the processes, the pain points and the real opportunities before choosing a tool.
  3. Build assistants and agentsTest a targeted case quickly in the existing environment, then add complexity only when it becomes necessary.
  4. GovernEmbed AI use in day-to-day routines, track its value, costs and risks, then improve or retire what no longer helps.

How I work: the Commando Sprint

A short engagement on a real episode of work. The client is not buying a number of assistants: they are buying the measured transformation of one activity.

  1. QualifyThe sponsor, the scope, the data, the test cases and the starting measurement.
  2. ObserveFollow a real business episode: its files, its decisions, its pain points and its workarounds.
  3. EliminateRemove, reduce or standardise what should be before bringing in AI.
  4. AugmentBuild the minimal solution with the users: assistant, agent, script or workflow.
  5. MeasureCompare before and after, note the limits, then decide: roll out, improve or stop.
We don’t just design an assistant. We design a new way of working.

What becomes decisive

Models can write, analyse, search and now act inside tools. But on their own they don’t know the right process, the authorised source, the business exception or the person in charge.

Producing a text or an analysis costs less and less. What matters becomes: choosing the right problem, providing the right context, fitting AI into a real process, checking its results and measuring what actually improves.

People
learn to delegate, describe and check.
Processes
bring in AI with steps, responsibilities and checkpoints.
Management
consolidates costs, risks and results to decide where to invest and what to stop.