What This Looks Like in Practice

Several of our clients receive more paper every morning than their teams can process. Broker submissions arrive as PDFs, scans and photographs, no two alike. A daily settlement package has to be read, reconciled and keyed in before the business day can start.

None of these clients had a software team, so the work scaled with headcount or it didn't scale.

Many pipelines with the same pattern. Documents arrive in a monitored mailbox or secure folder. A vision-capable model extracts the fields, including from scans and photographs with no selectable text. Internal rules are applied. What matters is surfaced in a clean interface, and what doesn't is filtered out before a person sees it.

Formatted responses are generated on the client's own letterhead. Everything runs inside the client's own cloud tenant.

Every proposal started from nothing. The answers existed. They sat in a project tracker nobody updated, a proposal library nobody opened, a search that returned everything and therefore nothing, and the personal folders of whoever wrote the last good one.

Estimators rebuilt the same arguments from memory, under deadline, across operating companies that could not see each other's work.

A three-part system. The first part reads an incoming RFQ and interprets what is actually being asked. The second searches the organization's proposal history, indexed by the question answered rather than by filename, so a buried paragraph from four years ago surfaces when it is relevant. The third assembles a sectioned draft into a working document, ready for a person to argue with.

The skeleton is generic, so it can be used again and again.

Mechanical and structural drawings hold an asset-heavy business's most valuable information, and they are the one document class that has stayed entirely manual. Labels, part numbers, revision marks and annotations sit as marks on a page.

Extraction means an engineer reading drawing after drawing, so it is done once, for one purpose, and never again.

A vision pipeline that treats a drawing as a readable document. It pulls from the existing repository, reads and classifies the labels, extracts annotations into structured records, and writes the result back where the business already works. There is no new system to log into, and no data leaves the security perimeter.

The underlying capability is already in production elsewhere. The same models that read scanned submissions also read drawings. What changes is the schema, not the architecture.

Several businesses acquired over a decade, each arriving with its own stack: a mix of ERPs, small-business accounting packages, SQL applications, an Access database and heavily programmed Excel workbooks. There was no shared part numbering, customer record, or definition of a quote.

Leadership could not answer "How are we doing?" without someone reconciling reports from every business by hand.

An ingestion and normalization layer across every entity, feeding one live dashboard. Awards and actuals are kept deliberately separate, because revenue is recognized at production rather than at award. Quote status is mapped to one vocabulary across systems that each named things differently. Historical data was pulled back years for each entity.

Then the part that matters: metrics the sales floor actually changes its behaviour for, such as open quotes, stale quotes, and the reasons a job was lost.

A relationship-driven business with hundreds of client organizations and a team small enough to count on two hands. Nobody could see when a client restructured, when a competitor launched a program, or when the person who signs the cheques changed jobs, until someone happened to mention it.

The information was public, but it was spread across more sources than anyone could read.

An OSINT (open-source intelligence) pipeline that monitors a defined universe of organizations and turns what it finds into a short weekly briefing: company news, leadership and learning-function hires, competitor program changes, and a shortlist of organizations worth approaching.

It was built once and pointed at three markets. Each region keeps the structure and drops the sections that don't apply.

Business development ran on who the team already knew. The CRM described current clients well and said nothing about the market beyond.

Prospecting meant someone finding an hour to think of names, which meant it happened when things were quiet. That is exactly when it is least useful.

An ideal-customer profile (ICP) derived from the client's own CRM, based on what their best accounts actually have in common rather than what anyone assumes. That profile defines a universe of comparable organizations, monitored continuously.

Each week the system proposes a small number of organizations outside the client list, with the specific signal that triggered each suggestion attached. A cooldown stops names from recurring.

An organization buys AI licences and the results underwhelm. The assumption is that the tool is weak. Almost always the tool is fine and the data is a mess.

A decade of documents with no consistent structure sits duplicated across drives, undated and contradicting itself, mixed in with meeting transcripts and interview notes nobody has read back. Point a capable assistant at that and it answers confidently from the wrong document.

A cleaning and restructuring layer that runs inside the client's own security tenant. It reads the existing document estate and the accumulated transcripts, resolves the entities within them (people, projects, clients, decisions, and how they relate), and writes a cleaned, deduplicated copy back into the same environment.

The licensed assistants already paid for are pointed at the clean copy. There is no custom retrieval system, no data leaves the tenant, and there is no new tool to learn. It is the same interface, suddenly answering correctly.

Staff were already using AI on personal accounts, with no controls, because the sanctioned tool was weaker than the free ones they had at home. Leadership had a budget question they could not answer: how many premium licences, for whom, and what comes after.

There was no acceptable use policy, no view of what was running on company machines, and no software engineering function to build whatever the answer turned out to be.

Three pieces that only make sense together.

Governance: an acceptable use policy that we drafted, the client's legal function reviewed and the General Counsel signed, paired with a data classification companion.

Licensing: real usage analyzed across more than 1,100 employees rather than estimated, producing a defensible seat count and an independent head-to-head evaluation run by the client's own heavy users.

Roadmap: a sequenced path from off-the-shelf agents, to low-code agent building, to fully scaled cloud implementation.

Adoption is almost never evenly distributed. Senior leaders cannot picture the strategic consequences, and the people doing the daily work cannot picture what it does for their job.

Train one group and not the other, and you get a strategy nobody executes, or enthusiasm with nowhere to point it.

For senior leadership: where AI changes the economics of the business, what it does not do, and which decisions need making now. These sessions run for executive teams, boards and committees, including in regulated and public-sector settings.

For emerging management and functional teams: practical literacy and change management, with hands-on work in the tools people already have, built around the workflows of the department in the room.

The two are connected. Leadership sessions surface the use cases, and team sessions turn them into habits.