For the pilot that never shipped
dotfun builds the version that survives: retrieval assistants, copilots, and document generation on your own data, with the evaluation and guardrails to turn it on.
.png)
Your pilot impressed the room.
Then it met the real data.
Generative AI is the easiest technology in years to demonstrate and one of the hardest to put in front of a customer. The gap is not intelligence. It is everything a demo skips.
The sample data was clean. Yours is not.
The prototype ran on a folder someone curated. Production runs on the shared drive:
Duplicates, three versions of the same policy, scans nobody converted to text.
Permissions were never in the design.
In the pilot, everyone could see everything. In production, an assistant that quotes the wrong document to the wrong employee is not a feature.
It is an incident.
Cost was invisible at ten users.
A long prompt over a long document is cheap once and a budget line at production volume.
Nobody modeled that before the rollout meeting.
Nobody agreed what "right" means.
There was no set of cases the output had to pass, so sign-off became opinion,
And the launch date kept moving.
A pilot answers one question, for one friendly user, on data you chose.
Production answers thousands, on data nobody cleaned. Our generative AI development services exist for that second job.
Five places a generative model earns its keep
The gen-AI lane of dotfun's AI Solutions practice, and the child of our AI software development work. If a model produces the thing your user reads, the build lives here. If it mostly acts inside your other tools, that is agent work or workflow automation.
RAG and knowledge assistants
Your policies, contracts, tickets, and product docs already contain the answer. The hard part is retrieval: putting the right passage in front of the model before it speaks (AWS explains the pattern well). We build the retrieval layer, the filtering that respects who may see what, and the citation trail that lets a person verify an answer.
Copilots and in-product AI
An assistant inside the product your team or customers already open every day. Drafting, next-step suggestions, answers grounded in the account in front of them. We build it into your interface, not beside it. A copilot people must remember to open is one nobody uses.
Document and content generation
Proposals, reports, briefs, and first drafts built from your templates, your tone, and your data, not a model's idea of your business. A person still approves it. The model removes the blank page, not the judgment.
Extraction and summarization
The quiet, high-value half of generative AI. Pull structured fields out of contracts, applications, forms, and long email threads, then write the result into the system that needs it.
Model selection and evaluation
Which model, at what cost per request, with what fallback when it fails. We test candidates against your actual cases, not a public benchmark, and design for the day your model gets deprecated or repriced.
How a generative AI build gets from prototype to production
The sequence is built backward from the failures above. Every stage exists because a pilot died there.
Before anything is built, we ask who reads the output and what they do next. A summary nobody acts on is a demo with a budget. If no decision depends on it, we say so, and you keep the money.
.png)
Most bad answers are not the model inventing things. They are the model never seeing the right passage. So we test retrieval against your real corpus first, in its messy state, before anyone argues about prompts.
.png)
You hand us the questions your team argues over and the answers that would embarrass you. Those become the evaluation set the build must clear before launch, and the regression check whenever a model or a document set changes.
.png)
The first version covers one document set, one team, one job, with human review on anything the system flags uncertain. We widen once the numbers hold.
.png)
the Iris by dotfun AI Maturity Assessment is free, runs 15 minutes, conversational, and sends a written action plan inside the hour. An optional 30-minute walkthrough follows. No credit card.
Start with the free assessment.
The fastest way in is the Iris by dotfun AI Maturity Assessment: a free, 15-minute conversational assessment that returns a written action plan within the hour, with an optional 30-minute walkthrough. No credit card. If the plan points to a real build, the senior-run assessment on this page picks up from there.
TAKE THE ASSESSMENT
Why a senior-run assessment beats the alternatives
The difference is who runs it, and what they can do about the answer.
Compared to an offshore development shop:
Most AI dev shops sell you a build, because a build is what they staff for. dotfun runs the assessment first and stays honest about whether you need a build at all, because senior operators run it, not a sales team. No junior handoff. No incentive to push the most expensive path.
Compared to a do-it-yourself AI platform:
Self-serve platforms hand you the tools and leave the hard part to you: deciding what to build and whether it pays back. That is the exact question the assessment answers. You get a diagnosis and a route, not a blank canvas.
Compared to doing nothing:
Waiting feels safe, and sometimes it is the right call. But "not yet" should be a decision you made on purpose, not one that happened because nobody looked.
What you do not have to worry about
An assessment should lower your risk, not raise it.
We do not need you to build something to make the assessment worth it. If the honest read is "fix the data first" or "not yet," that is the recommendation you get. A diagnosis you can trust is the point.
The assessment routes you to the best path, not to a dotfun contract. The plan is yours. Build it with us, take it in-house, or wait until the timing is right.
Looking at where your knowledge lives does not mean handing over control of it. We assess your data and systems to understand readiness, and your information stays yours.
The people assessing your business are the same senior team that would build the result, with 30-plus combined years in B2B tech. No junior analyst. No handoff.
Built by the team that builds both paths.
workflow automations and custom agents — and it builds Iris. That is why the route can be neutral: we deliver both the custom build and the product
And whatever gets built has to move the business. An automation that never touches the funnel is a cost, not a return. Our growth side, Feel Good, is where an AI investment shows up as pipeline, so what we recommend ships with the tracking and growth motion behind it.
Frequently asked questions
If a model produces the thing your user reads, an answer, a draft, a summary, a document, the build lives here. If it mostly acts inside your other tools, that is agent work or workflow automation. We scope which one your problem needs before anything gets built.
Usually four reasons: the sample data was curated and yours is not, permissions were never in the design, cost was invisible at ten users, and nobody agreed what right means. Our build sequence is written backward from those four failures.
Permissions are part of the retrieval design, not a layer added afterwards. The filtering respects who may see what, and every answer carries a citation trail so a person can verify it against the source document.
You hand us the questions your team argues over and the answers that would embarrass you. Those become the pass-or-fail evaluation set the build has to clear before launch, and the regression check whenever a model or a document set changes.
Whichever one clears your cases at a cost per request you can live with. We test candidates against your actual data rather than a public benchmark, and we design for the day your model gets deprecated or repriced.
Narrow on purpose: one document set, one team, one job, with human review on anything the system flags uncertain. We widen once the numbers hold. Starting narrow is how a build survives contact with real data.
Built for teams weighing an
AI investment
An AI readiness assessment fits the leaders who suspect AI could help but do not want to burn a quarter guessing.
Founders and operators with a mandate and no map
You have been told to "do something with AI," and you want a real answer about where it pays back before you commit budget. The assessment gives you that answer and a plan to act on.
Teams that have been pitched a build
You have a proposal on the table for a custom AI project and you want a neutral read before you sign. The assessment tells you whether the build is right, or whether a packaged product does the job for less.
Leaders who want senior guidance on AI
If you want a senior partner to guide your AI decisions, start here with a low-commitment diagnosis, then move into ongoing AI consulting if the roadmap warrants it.
