Search for AI deployment and you'll mostly learn how to put a model on a server. That's the right answer to a narrow question. Most organisations we work with don't train models: they buy them, as chat assistants, copilots and features inside software they already pay for. For them, the hard part is getting a whole organisation to use AI well, and that's where most of the money goes missing.
What is AI deployment?
AI deployment is the work of moving AI from a test into everyday use: in the technical sense, putting a trained model into a live system where it handles real requests, and in the organisational sense, getting AI into the way people actually work, with owners, guardrails and a measured result.
This guide covers both. It's for leaders who own an AI rollout in organisations of roughly 20 to 1,000 people, and for anyone deciding who should help them. The first section answers the technical question briefly. The rest is about the part the technical guides leave out.
How do you deploy an AI model?
In engineering, deployment means taking a trained model out of development and putting it where software and people can call on it. The cloud platforms offer a few ways to do that. The right one depends on how quickly an answer has to come back.
| Method | How it works | Use it when |
|---|---|---|
| Real-time (online) | The model sits behind an endpoint and answers each request as it arrives | Answers are needed in moments, such as a chat reply |
| Batch | The model scores a large dataset in one job, with no endpoint left running | Results can wait, such as scoring every account overnight |
| Asynchronous | Requests join a queue and are processed as capacity allows | Inputs are large or slow to process |
| Serverless | The platform runs and scales the model for you | Traffic comes in bursts, with idle periods between |
| On-device (edge) | The model runs on the phone, laptop or machine itself | The model has to work where the device is |
Getting a model live safely is its own discipline. Three patterns do most of the work:
- Shadow testing. The new model receives a copy of live requests, but only the current model's answers go back to users. You compare the two before anything changes.
- Canary release. A small share of traffic goes to the new model for a set period. If nothing trips the alarms, the rest follows.
- Blue/green with rollback. The new version runs alongside the old one and traffic shifts across. Automatic rollback switches it back if something goes wrong.
None of this is the hard part. A 2015 paper by Google engineers, Hidden Technical Debt in Machine Learning Systems, found that "only a small fraction of real-world ML systems is composed of the ML code", and that "the required surrounding infrastructure is vast and complex". And many organisations of 20 to 1,000 people never do any of it themselves. They use someone else's model through a product or an API, and the vendor does the model deployment. In McKinsey's 2026 survey, chatbots were the most widely scaled kind of AI tool: 47% of respondents said their organisations were scaling them across the enterprise.
For those organisations, deployment means something bigger.
What is the 10/20/70 rule for AI?
The 10/20/70 rule is BCG's guide to where AI effort should go: 10% to algorithms, 20% to technology and data, and 70% to people and processes.
It comes from BCG's 2024 study Where's the Value in AI?, a survey of 1,000 senior executives in more than 20 sectors. In BCG's words, leaders "follow the rule of putting 10% of their resources into algorithms, 20% into technology and data, and 70% in people and processes." The same survey found that about 70% of the problems companies hit when implementing AI came from people and processes, 20% from technology and only 10% from the algorithms. Just 26% of companies had built what it takes to move beyond proofs of concept and produce real value.
Everything in the section above sits in the 10% and the 20%. It matters, and it's the part a vendor or an engineer can do for you. The 70% is the part only your organisation can do: deciding what AI is for, redesigning the work, training people on their own tasks, naming owners and measuring the result. The rest of this guide is about the 70%.
Why do most AI deployments stall after the pilot?
Pilots are easy to start and hard to finish. The best evidence says most never return anything.
In July 2025, MIT's Project NANDA published The GenAI Divide. It drew on a review of more than 300 public AI initiatives, interviews with people at 52 organisations and a survey of 153 senior leaders. Its headline finding: despite $30 to 40 billion of enterprise investment in generative AI, "95% of organizations are getting zero return." For enterprise-grade tools, custom or bought from a vendor, 60% of organisations evaluated them, 20% reached a pilot and 5% reached production. The authors put most failures down to "brittle workflows, lack of contextual learning, and misalignment with day-to-day operations." They also call their figures "directionally accurate", based on interviews rather than company reporting. Treat the 95% as a strong signal, not a precise count.
McKinsey's 2026 survey shows where the gap opens. It heard from 1,719 people in 97 countries, and 80% of them said AI had improved their own productivity. Only 37% said AI had made any contribution to their organisation's EBIT, about the same share as a year earlier. People are faster. Their organisations, mostly, aren't better off.
Three causes come up again and again.
AI speeds up work that wasn't holding anything back
An organisation moves only as fast as its slowest step. Speed up any other step and the extra work piles up in front of the slow one. That's how AI makes individuals faster without making the organisation any better, and why so much AI output lands on someone else's desk to fix. The first lesson of our free course, Why most AI rollouts return nothing, walks through it.
Nobody owns it after the pilot
A pilot has a sponsor and an enthusiastic team. Production needs an owner, a backup, an off switch, a review date and a budget line. When the pilot team moves on, those are usually missing. MIT's survey found the most common barrier to scaling was resistance to adopting new tools, followed by doubts about the quality of the output. A better model fixes neither on its own.
People hide how they use it
MIT found that only 40% of companies had bought an official subscription to a large language model, while workers at over 90% of the companies it surveyed used personal AI tools regularly. When people worry about what disclosure will cost them, the organisation can't see what works, so it can't spread it. Our guide to shadow AI covers how to bring that use into the open.
The organisations getting a return do the opposite. McKinsey counts about 6% of its respondents as "AI high performers": they attribute at least 5% of EBIT to AI and call its impact significant. Nearly three-quarters of them had fundamentally redesigned workflows because of AI, against a quarter of everyone else. They were also twice as likely to say their senior leaders show commitment to AI, and that they have defined ways to measure its impact.
Deployment is a loop, not a launch
The technical kind of deployment has an end: the model is live. The organisational kind doesn't. Fixing one bottleneck exposes the next, so deploying AI across a company works as a cycle you run every quarter, not a project you finish.
That cycle is the core of the Deployment Framework, the free framework we use with clients. In brief:
- The aim: one page that says what AI is for, with a goal you can measure and a date, where freed-up time goes, a spending ceiling and who's responsible.
- The promises: telling us how you use AI is safe, nobody is punished for improving their own work, and if roles are ever cut, it happens through a separate, announced process that offers redeployment first.
- Four moves: find the bottleneck, improve the work without spending, build only what improvement can't fix, and own what runs.
- The guardrails: approved tools, an AI list, a data map, an output standard, spending limits, and a blame-free way to stop, tell, fix and learn when something goes wrong.
- The rhythm: a short weekly check in every team, a quarterly leadership review and a yearly review of the guardrails.
The framework page explains each part in full. The stages below put it in order.
What are the stages of deploying AI in a company?
There are eight stages. Each one produces what the next one needs.
01 · Set the aim and make the promises
Leaders agree a one-page aim and announce it before anything else happens: a goal you can measure, with a date, where freed-up time will go, a spending ceiling with a small budget for experiments, and a sponsor plus one named person who runs the work. Then they sign the promises, because nothing else works if people hide how they use AI. The four promises sets them out. The aim is what every later stage is measured against.
02 · Put the guardrails in place
Six come first: approved tools (including the AI features vendors switch on inside software you already have), the AI list, a data map of which data can go into which tools, an output standard, spending limits, and a blame-free path for when something goes wrong. Guardrails work when they're built into the tools rather than buried in a policy document. Your starter guardrails gives you the set.
03 · See what's already running
Your organisation is already using AI, approved or not. Invite everyone to add their use to the AI list, under the promise that telling you is safe, and check what vendors have switched on. The free Shadow AI Audit builds a first version in about five minutes. Your AI list turns it into a register with an owner on every row.
04 · Find the bottleneck
Trace the work back from the customer and look for where it waits, piles up or gets redone. Name the one step that most limits the aim and give it a lead. Before you touch anything, write down what you expect to change, how long you'll measure, and the dip you expect while people learn. Naming one bottleneck shows how to put a value on it.
05 · Improve the work before you build
Redesign the work around the bottleneck with the tools you already have. Train people on their own work, in protected hours, managers first. Pace everything else to what the bottleneck can absorb. This stage spends no new money, and redesigned workflows are one of the clearest differences McKinsey found between its high performers and everyone else. Redesign before you automate explains how.
06 · Build only what improvement can't fix, and prove it
If the bottleneck didn't move and closing the gap is worth paying for, write a one-page case: the value, the full running cost, an owner and a backup, a data check, a stop-loss and kill criteria. Then prove it in two stages: first in a sandbox that can't reach live systems, then in a limited live trial with real users in the systems they already use. This is where the technical patterns belong. A shadow test or a canary release is how engineers run that limited live trial. Two-stage proof covers both stages.
07 · Own what runs
Everything shared or high-stakes gets an owner and a backup who can stop it, monitoring, an off switch and a tested undo, a plan for when the vendor changes it, and a review date. Owners and backups sets them up.
08 · Measure at day 90, then find the next bottleneck
Score the bottleneck against your prediction. Check where freed-up time actually went, because it only counts once it has gone where the aim said. Keep, fix or retire what you built, and name the next bottleneck. Where freed time actually went shows how to trace it.
Our free course, Your first 90 days with AI, takes a leadership team through all eight stages, one module each.
AI deployment checklist
Use it before you call any deployment done.
Before you start
- A one-page aim, with a measurable goal, a date and where freed-up time goes
- The promises, signed by leadership and announced to everyone
- A sponsor, and one named person who runs the work
- A spending ceiling and a budget for experiments
Before anything goes live
- An approved tools list, including AI features switched on inside software you already have
- An AI list that matches what people actually use
- A data map: which data may go into which tools
- An output standard: AI-assisted work is labelled, owned, kept short and checked before it's passed on
- One named bottleneck, with a lead and a written prediction
- For anything you build: a one-page case, a sandbox test and a limited live trial
- For anything high-stakes: evidence from someone who didn't build it
Once it's running
- An owner, and a backup who can stop it
- Monitoring, an off switch and a tested undo
- A plan for when the vendor changes the model or the terms
- A review date in the calendar
- A record of where freed-up time went
- A weekly team check, a quarterly leadership review and a yearly review of the guardrails
To see where you stand, the free AI Readiness Assessment scores your organisation from 0 to 100 across six areas, including strategy, people, governance and tooling, and gives you a next move for each.
AI deployment examples
Four public examples. Each one carries a lesson well beyond its own industry.
The UK government: 20,000 Copilot licences, and time nobody could trace
From 30 September to 31 December 2024, the UK's Government Digital Service gave Microsoft 365 Copilot licences to 20,000 civil servants. Earlier trials had been limited to 300 licences at most. Participants estimated they saved an average of 26 minutes a day, though 17% noticed no clear saving. The report is candid about what it couldn't show: "it was not possible to identify how time saved was spent." Driving adoption "required substantial effort", and support fell to individual departments. It also warned that because Copilot searches with each user's access, it "may highlight issues when users have access to files they shouldn't."
The lesson: time saved isn't value until you can say where it went. Decide that before the rollout, and fix file permissions before an assistant can search them.
Morgan Stanley: tested before it was rolled out
Morgan Stanley's wealth management business built an assistant on OpenAI's models to answer financial advisers' questions from the firm's own documents. According to OpenAI's case study, the firm set up "an evaluation (eval) framework to test every AI use case before deployment", with advisers and engineers grading the answers, and it runs daily tests against a set of sample questions. Over 98% of its adviser teams now use the assistant. For a second tool that summarises client meetings, advisers review and adjust the output before it's final.
The lesson: prove it on your own work before you scale it, keep testing after launch, and keep a person checking anything that reaches a client.
Klarna: customer service at scale, then a human option
In February 2024, Klarna said its AI assistant had handled 2.3 million conversations in its first month, two-thirds of its customer service chats. It said the assistant was doing the equivalent work of 700 full-time agents and that customers were resolving their issues in under 2 minutes rather than 11. In May 2025, according to reporting by Entrepreneur on an interview with Bloomberg, Klarna's chief executive said its AI customer service chatbots, though cheaper than human staff, had resulted in "lower quality" output, that customers should know "there will always be a human if you want", and that investing in human support was "the way of the future" for the company.
The lesson: measure what customers experience, not only the cost per conversation, and keep a route to a person for anything that matters to them.
Air Canada: the chatbot speaks for the company
In February 2024, a Canadian tribunal ruled against Air Canada after the chatbot on its website told a customer they could apply for a bereavement fare after travelling. The airline's own bereavement page said the policy didn't apply once travel was complete. Air Canada argued, in effect, that the chatbot was responsible for its own actions. The tribunal called that "a remarkable submission" and found the airline "did not take reasonable care to ensure its chatbot was accurate": "It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot."
The lesson: whatever your AI tells customers, your organisation told them. Give every customer-facing use an owner, a check and a way to correct it.
AI agent deployment: what changes
An agent doesn't just answer questions. It takes actions in your systems, which raises the stakes at every stage above.
In June 2025, Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, "due to escalating costs, unclear business value or inadequate risk controls." It also warned about "agent washing", where vendors rebrand chatbots and assistants as agents, and estimated that only about 130 of the thousands of agentic AI vendors are real. Adoption is uneven, too. In McKinsey's 2026 survey, 40% of respondents from organisations with more than $1 billion in revenue said they were scaling AI agents, up from 27% a year earlier. At smaller organisations the figure stayed flat at 22%.
Our framework adds rules for agents on top of the starting guardrails. Each agent gets its own login with limited reach, never a borrowed human account. It stops itself when its output looks wrong, and anyone can stop it. New builds can't reach live systems until they're promoted to a limited trial, and they only get full reach once they're kept. Rules for agents covers how to set them.
How to choose an AI deployment partner
Many organisations get help with some of this. MIT's study found that AI built through external partnerships reached deployment about 67% of the time, against about 33% for tools built internally. The authors caution that the figures are self-reported and that the gap may reflect the organisations as much as the approach. Building is also getting easier: in McKinsey's 2026 survey, 32% of respondents said their organisations had decided against buying at least one software product or feature because they could build it with agentic coding tools. McKinsey's commentary on the survey says leading organisations now engage partners "from a position of agency", deciding deliberately where to buy and where to build.
There are four kinds of help.
The AI labs' own forward-deployed teams
Some model makers now have engineers who work inside customer organisations. OpenAI describes its forward deployed engineers as leading "complex end-to-end deployments of frontier models in production alongside our most strategic customers." Nobody knows the models better. Weigh two things: they deploy their own company's models, and they work with the customers the lab counts as its most strategic.
Large consultancies
Global firms bring people in volume, programmes that span many countries and functions, and change teams alongside the technologists. They suit large, multi-year transformations. Ask who will do the work week to week, how they'll hand it over, and whether the plan starts from your bottleneck or from a standard programme.
Boutique deployment firms
Small specialist firms concentrate on the organisational side: the aim, the guardrails, the bottleneck and ownership, often for organisations too small for a large consultancy's programme. Ask how much they build themselves, and who they bring in when a build needs serious engineering. We're one of these, so weigh this paragraph with that in mind.
In-house teams
Your own people know the work, and the capability stays when the project ends. The risk is that a capable technical team covers the 10% and the 20%, and nobody owns the 70%. In MIT's sample, internal builds reached deployment about half as often as external partnerships.
Whoever you choose, ask them:
- Will you start from our bottleneck, or from a product you already sell?
- What will you measure, against what prediction, and by when?
- Who owns what you build after you leave, and how is it switched off?
- How will you handle the people side: the promises, the training and any changes to roles?
- Will you work inside our guardrails and our data map?
- What will our own team be able to do without you at the end?
If you'd like help, our AI implementation service runs the framework with your leadership team. A consultation is the place to start.
Sources
- The GenAI Divide: State of AI in Business 2025, MIT NANDA, July 2025 (copy hosted by MLQ.ai): the 95% finding, the pilot-to-production figures, shadow AI use, barriers to scale, partnership versus internal builds and time to implement.
- The state of AI in 2026: On the road to ROI, McKinsey, August 2026: individual productivity against EBIT impact, high performers, chatbots and agents at scale, and building instead of buying.
- AI Adoption in 2024: 74% of Companies Struggle to Achieve and Scale Value, BCG, October 2024: the 10/20/70 rule and where implementation problems come from.
- Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, Gartner, June 2025: the agentic AI prediction and agent washing.
- Microsoft 365 Copilot Experiment: Cross-Government Findings Report, Government Digital Service, June 2025: the UK government example.
- Klarna AI assistant handles two-thirds of customer service chats in its first month, Klarna, February 2024: the launch figures.
- Air Canada chatbot decision (2024 BCCRT 149), Civil Resolution Tribunal of British Columbia, February 2024: the ruling quoted in the Air Canada example.
- Deploy models for inference, Batch transform, Deployment guardrails and Shadow tests, Amazon SageMaker documentation: deployment methods and safe-release patterns.
- Endpoints for inference, Azure Machine Learning documentation, August 2026: online and batch endpoints.
- LiteRT, Google AI Edge: on-device deployment.
- Hidden Technical Debt in Machine Learning Systems, Sculley and others, NIPS 2015: how little of a production ML system is the model.
- Morgan Stanley uses AI evals to shape the future of financial services, OpenAI: the Morgan Stanley example (a vendor's case study).
- Forward Deployed Engineer (FDE), San Francisco, OpenAI careers, October 2026: how OpenAI describes its forward-deployed role.
- Klarna CEO Reverses Course By Hiring More Humans, Not AI, Entrepreneur, May 2025 (secondary, reporting a Bloomberg interview): Klarna's turn back to human support.