These are the questions our clients ask most before a project starts, answered as we would on a first call. They cover what we do, what it costs, who does the work and what happens next, and each answer shows its sources so you can check it. Together they give you enough to decide whether to book a call.
1 of 7
What AI development services do you offer?
Winder.AI answers
Applied AI research and development: bespoke applications, built to run in production.
Clients come to us when a problem needs automating, often using their own proprietary data. We build bespoke systems using machine learning, data science, reinforcement learning, LLMs and agents. Our work runs from pure research to production applications. We have done this for over a decade1.
We build AI clones that assist your experts, agent workflows that run a process end to end, and digital teams of agents that work alongside your staff, so your best people’s knowledge reaches every customer and colleague at once. AI Phil on this page is one such clone.
Temple University: a legal AI assistant, nearly 80% accurate at finding the right legal text in testing5.
For teams who want to deploy AI faster or have many models in production.
We review how your models reach production, then design and build the deployment pipelines and monitoring they need, so each release is quicker and you know when a model stops performing.
Tractable: a return on investment greater than 100x6.
Not sure which service fits your problem? An AI readiness assessment takes two to four weeks and ends in a written report and an executive readout7. AI readiness assessment.
Should we build custom AI or use managed services?
Winder.AI answers
Build custom when the use case is unique and here to stay.
AI has made a fully custom system practical to build, and a managed cloud runtime keeps it as reliable as a managed service. So the test is the use case: how unique it is, how long it must last, how much it will change and how many people will use it.
Managed AI services
For generic work, few users or a short life.
You configure a service such as Copilot Studio and pay for what you use. It is the quickest start, and you work within the vendor’s limits: in Copilot Studio a prompt has 100 seconds to run8.
Open source, run yourself
For the middle ground.
An open-source system on your own infrastructure gives you more control than a managed service, with less to build than a custom system. You take on running it. Where data cannot leave your network, open-source models run on your hardware9.
Custom build
For unique work, many users or a long life.
We write the system for your use case and run it on a managed runtime in your cloud, so it changes when your needs do and the platform handles scaling and uptime. On one report-writing job, a firm writing thirty or more reports a year was better off building8.
With a custom build
Where it runs
In your own cloud tenant or on-premise10. We deliver across AWS, Azure, GCP and on-prem Kubernetes, including air-gapped environments11.
Your data stays yours
We sign data processing agreements, never use your data to train AI models, and can host workflows on UK infrastructure if required12.
No lock-in
We are model-agnostic across OpenAI, Anthropic, Google, Llama and Qwen13. On reinforcement learning projects you own the environment, the trained policy and the code14. You own the IP for the custom AI we build for you13.
Shown work
Managed or custom on Microsoft's stack: what one document costs
A firm asked what it would cost to generate a 236-page report with AI on Microsoft’s stack. The same tokens, priced on the same day, cost very different amounts depending on where you build it8.
A follow-upWe are already on Microsoft. Would it not just sit in Copilot?
Winder.AI answers
Sometimes it should. Copilot’s Agent Builder comes with the Microsoft 365 Copilot licence, but Microsoft says it does not support complex workflows and integrations8. Copilot Studio can, and that is where the bill changes: one long report cost $725 there and about $4 through the model API, both on Microsoft’s own prices8.
What you give up by leaving Copilot is identity and governance: it answers as the signed-in user and respects sensitivity labels, and rebuilding that is the expensive part8. On that job, a firm writing thirty or more of these a year is better off building8. We sell that build, so check the working in the article.
Production AI in finance, travel, energy, technology and the public sector, with the results published.
We have delivered production machine learning since 201315. Every project below has a published case study, with the client named wherever the client allows it. Pick your sector to see the work closest to yours.
>100x
Tractable: return on investment from our MLOps consulting6.
500k+
Stability AI: Stable Audio generations in its first two months2. Stable Audio was named one of TIME’s Best Inventions of 202316.
80%
Grid.AI: lower cost of training a 1000-node model17.
9.4
Average of seven scored client responses. Interos, Apartment List and Expanso each gave 10 out of 1015.
“I would recommend Winder.AI because they are experts with real-world experience, led by Phil Winder, who is well-respected in the industry. They are quick to respond, quick to scale up and they deliver when you need them to.”
Hunter Powers, Vice President of Machine Learning, Interos25You can see our work at
Probably, or next to it. Beyond those five sectors, the published work includes hotel pricing for Duetto18, a report for the UK communications regulator Ofcom21, legal research at Temple University5 and process control at a paper mill for CMPC22.
Pick your sector in the list above to see the closest work. If yours is not there, ask on the call which project is nearest to it.
Every project gets its own plan, guided by you and your team.
We start from where you want to get to and write a plan to get there: the stages, what each delivers and when. You and your team guide the plan and the development throughout. We can work the way your team works, in its tools and routines, or independently, and we fit the process and current tooling to the work in hand.
You bring
The problem, your data and systems, and the people who know how the work is done today.
We bring
The plan, the engineers who design, build and measure the system, and the tooling the work calls for.
The stages, adapted to what you need
1Discovery
Two to four weeks
We audit your data, your process and your goals. A consulting engagement ends with a report. A development project ends with a plan, and sometimes a very limited prototype that shows how the whole process will work.
2Proof of concept
Two to six weeks
We build a working version against your real data and measure it. A pure proof of concept ends with a very limited prototype that you can look at. A development project ends with something you or your users can use.
3Production build
Two to six months
We build the production system, with pipelines, integration with your stack, audit trails and monitoring, rolled out in stages so the impact is measured26.
We monitor the system, retrain it and respond to incidents26. Or your team takes over: a production build includes handover27, and at Apartment List we trained the internal team to run it themselves19.
In practice
How the plan differed at Duetto and Apartment List
Duetto was a research question, and its plan had four phases over five months: discovery, a behavioural cloning baseline, offline reinforcement learning experiments and an evaluation method18. Apartment List was a platform, and its plan had three: discovery and assessment, architecture design, and a roadmap with enablement19.
Then you find out early, and we change the plan with you. A proof of concept tests the idea on your real data before any production build, and you get its verdict in writing14.
An automation audit is a standalone engagement with no further obligation12.
Dr Phil Winder, on every project. A small team of engineers joins when the work needs more hands.
Phil is involved in every project we take on. When the work needs more people, he brings in a small team of engineers. We make heavy use of AI automation in our own work to give you the best value for money, which is why most projects need one to four engineers.
Phil holds a PhD in electronic engineering, wrote O’Reilly’s Reinforcement Learning, and has trained thousands of data scientists29.
“The most unique part about working with Winder.AI was Phil's involvement, keeping his pulse on things, and helping us with larger strategy and planning.”
Hunter Powers, Vice President of Machine Learning, Interos20
How big the team is
There are rarely more than ten of us in action at once across everything, and the people who scope a project are the people who build it15. There is no offshore handover and no junior squad behind a senior pitch11.
Where they are
Winder.AI is registered in the UK. Our engineers are in Europe and the USA, working remotely1.
How we staff a project
Other engineers join as associates, so we bring in the right talent for each project1.
Where we are the wrong fit
The team is small and senior, so we are the wrong shape for staff augmentation28, and for anything needing a 400-person bench next month15.
The engineers who work alongside Phil are on our team page: Meet the team.
AI development on time and materials runs at £150 to £350 per hour, depending on the commitment30. A fixed price runs from roughly £30k for a focused project to £100k to £250k for a multi-phase production system30.
Hourly
Time and materials, for advisory work and development.
£150 to £350 per hour, depending on the commitment30
The vast majority of our development projects are structured this way31.
Fixed price
A well-scoped project with agreed deliverables.
£30k to £250k, from a focused project to a multi-phase production system30
These are indicative prices from real projects, not a quote for yours31.
Examples
Indicative pricing
A low-commitment start
For a team that wants advice before it commits to a project.
Ten hours of Phil’s time to test an idea or review a plan: £3,500
Hourly, at £350 per hour.
A focused project
For a team with one well-defined problem.
Two to six weeks, ending in something you or your users can use: from £30k
Fixed price.
A production AI product
For a product or system the business will depend on.
Two to six months, delivered in phases: £100k to £250k
Fixed price.
An ongoing team
For a company that needs engineers month after month.
Engineers on a longer commitment: from £30k a month
Hourly.
These examples are worked from the rates above. We quote each project on its scope.
Beyond the price
Running costs and ongoing support
After the build
That depends on what is built. A system at massive scale, running on tens of GPU machines, gets expensive quickly. A highly optimised system built on off-the-shelf AI can cost very little to run. Consulting and some data science projects have no running cost at all, because the result is a research report.
Ongoing support
That depends on how much you want us to do. Light monitoring of a system we built is about £3k a month. Retraining and updating the models, and responding to new feature requests, is about £15k a month.
Start with the ranges above: roughly £30k for a focused fixed-price project, up to £100k to £250k for a multi-phase production system30. Where yours lands depends on the scope and how ready your data is32.
It starts with a short brief, then a 45-minute call with Phil.
Tell us about the project in the form below, as it is: half-formed is fine. Phil reads every brief and replies to arrange a call. Our calendar fills four to eight weeks ahead, so get in touch early30.
Tell Phil about your project
Send a short brief in the form. Phil replies to arrange a 45-minute call, and you will leave that call knowing whether we can help.
Your brief
Write as much or as little as you like. We reply to every message.
From brief to kick-off
1
A short brief. You tell us about the project in the form. Phil reads it and replies to arrange a call.
2
A first call. Forty-five minutes with Phil. We understand what you are trying to achieve and check we are a good fit31.
3
A second call. We agree how the work is priced and structured31.
4
A written proposal. Scope, people, price and timeline, in writing31.
5
Contract and kick-off. Typically two to four weeks after the first call: one to two weeks of scoping, then one to two of contracting. For urgent work we have started inside a week30.
Ask AI Phil
Pl. 195 · LLM Inference
Pl. 197 · Harness Engineering
Pl. 198 · Demand forecasting
