AI built for engineers.

The report is the deliverable — and the bottleneck.

Engineering consultancies run on reports — geotechnical, civil, structural, traffic and more. Writing them is slow, repetitive work that most engineers don’t especially enjoy.

AI is not coming to this work; it is already here. Junior engineers are using AI to help them write, whether their firms encourage it or not. The trouble is they are using general-purpose chatbots that are not built for engineering. Those tools invent numbers and facts, produce pages of unnecessary text, and miss the context an engineering report needs. The result lands on a senior engineer’s desk as text that has to be heavily checked and rewritten — a real liability, and a real time sink.

How Harvey proved this works — in law.

Harvey is an AI assistant built specifically for lawyers. Law is one of the hardest industries to bring AI into: the work is high-stakes and confidential, and the profession is conservative and risk-averse — a wrong or invented answer can be a serious liability. Harvey won that industry over by building trust rather than novelty: it works from a firm’s own documents, shows its sources, and is tuned to legal work instead of being a general chatbot.

It also started simply — leading general-purpose AI models combined with document retrieval, with no expensive custom training up front. That was enough to prove the concept and win over major firms; only later, as it grew, did it invest in custom-trained models. Today Harvey is used across many of the world’s largest law firms, including firm-wide rollouts of thousands of lawyers. It has reportedly reached around US$300 million in annual recurring revenue, was valued at roughly US$11 billion in early 2026, and is expanding into neighbouring professions such as tax and accounting.

And Harvey is not a one-off. The same pattern — an AI built for one profession, earning trust where general tools can’t — is repeating in other document-heavy fields:

• Law — Legora, a direct Harvey rival, recently valued at around US$5.5 billion.
• Accounting — Basis, around US$1.15 billion, already used by roughly a third of the top-25 US accounting firms.
• Medicine — Abridge, around US$5.3 billion, for clinical documentation.

Strong pull, the right gap, a proven model.

Engineers are already adopting AI quickly, but with the wrong tools — most knowledge workers now use AI at work, the majority on their own general-purpose apps with little firm oversight, and that pull is sharpened by flat fees and a deepening senior-engineer shortage. On the supply side, a few early companies build pieces of this — some answer questions over project documents, others draft a single type of report — but none yet combine project knowledge, question-answering and multi-discipline report writing with the trust controls engineering work demands. And the model is already proven next door: vertical AI for law, accounting and medicine has produced several multi-billion-dollar companies in barely two years.

Answer the questions. Write the reports.

Gesher is an AI built for engineers. From a project’s own documents it does two connected things — answers questions about the project, and helps write the reports — both without inventing anything.

• You give it the project’s documents — the proposal, the site investigation, lab results, survey, and so on.
• It reads them into one verified record of the project’s facts, and remembers where every fact came from.
• Ask it anything about the project — “what does the geotechnical report say about groundwater?” — and it answers from those documents and shows you the source. (The same kind of research assistant Harvey gives lawyers.)
• If two documents disagree, it does not guess. It flags a conflict, records neither value, and waits for a human engineer to decide which is correct. The AI never picks the winner.
• When you’re ready, it helps write the report and exports a Word document in which every passage of AI-written prose is clearly marked for engineering review.

The trust general tools can’t give.

Most engineering consultancies are cautious with new technology. Today, the only AI most of their staff can use is Microsoft Copilot, restricted to whatever underlying models the firm has allowed. Those are general-purpose tools: they hallucinate, they aren’t built for engineering work, and they give the firm little control over how confidential project data is handled.

To market quickly. Improve as we grow.

• Start simple and proven — the best available leading AI models plus document retrieval, with no expensive custom training up front. We’re not tied to any one AI provider; we use whichever model is best for the job.
• Go to market, earn revenue, reinvest — once it’s generating income, enhance the system as funding allows (for example, training a custom model to lower running costs and better match each firm’s voice). An upgrade revenue pays for, not something needed before launch.
• Built to scale — the architecture is designed so new report types and engineering disciplines can be added over time, rather than rebuilt each time.

It’s real. I want a partner to build it.

• The need is real — supported by market research and by early conversations with engineers.
• The approach is sound — Harvey has shown this model works in a comparably resistant, risk-averse profession.
• There is a clear gap — our research found no one has built the complete version for engineering.
• It is built and in testing — an early working prototype exists, and every test to date has passed.