AI readiness is a capability question. A business is ready for AI when it has the people to connect the tools to its data and systems, check what they produce, keep them secure and run them after launch. In our experience, most businesses buy the tools first and discover those gaps later, when the pilot has to work in the real operation.
UK figures point the same way. In Skills England's research on AI upskilling (July 2026), based on a UK-wide employer survey, workshops and case studies, more than 44% of organisations used AI tools daily, but only 1% had reached the stage of scaling AI, and technical AI skills were the gap employers reported most often (67%). AI agents raise the stakes: an agent that updates records or sends messages is only as reliable as the data, connections and checks around it. Our first article in this series covered how AI is changing what companies offshore; this one covers the support structure.
Key takeaway
We group the support structure around AI into six capabilities: data engineering, data quality, integration, quality assurance, security and AI operations. Most of that work continues after launch, so each needs a named owner and a budget. Buy specialist help for one-off design work, build or blend the continuing work, and keep ownership of your data, your decisions and the knowledge of how your systems behave. Where the continuing work is specialist and slow to hire in the UK, an embedded offshore team is one way to build it under your own management.
Why is buying an AI tool only the start of AI readiness?
Rolling out an assistant such as Claude or Microsoft Copilot is the quick part. Connecting it to the finance system, the CRM and the document store, deciding what it may see, checking what it produces and keeping those connections working as the systems change is what decides whether it pays back. Finance, sales, HR and operational systems often do not talk to each other, and connecting them can create value of its own.
The evidence on stalled projects points to the same gaps. In a Wakefield Research survey for GFT Technologies of 945 CIOs and CTOs at companies with revenue of at least $500 million (September 2026), 84% said limitations in their legacy systems had caused their organisation to cancel an AI pilot or project. Agents widen the gap: in Deloitte's survey of 501 US leaders at organisations already piloting or using AI agents (August 2026), only 5% said their business processes were highly prepared for them. In our experience, projects stall when the capability around the tool is missing: data nobody has prepared, systems nobody has connected and controls nobody is running.
What capabilities does AI need around it?
Each of the six shows up in a recognisable way when it is missing.
| Capability | What the work involves | How you know it is missing |
|---|---|---|
| Data engineering | Building and monitoring the pipelines that move data from your business systems to where AI tools can use it | Each new use starts with manual data work, and answers go stale when a source changes |
| Data quality | Validating, cleansing and enriching records before AI reads or writes them, and fixing the causes of bad data at source | The tool is confidently wrong about customers, products or prices, and staff stop trusting it |
| Integration | Connecting AI tools to your systems of record with the right permissions, and keeping the connections working through upgrades | People copy and paste between the AI tool and your core systems |
| Quality assurance | Testing software and AI-enabled workflows before and after each change, and sampling AI output against an agreed standard | Customers or auditors find the errors first |
| Security | Controlling what AI tools can see and do, protecting data going in and out, and planning for attacks specific to AI | Nobody can say exactly which data the AI can reach |
| AI operations | Monitoring accuracy and agent activity, handling incidents, applying vendor and model updates, managing running costs and supporting users | After launch, issues queue up because nobody owns the tool |
In the businesses we work with, data work is usually the largest of the six, and agents make it larger. In Deloitte's survey, 72% of leaders said they lacked unified, accessible data for agents and 67% said integration was too costly and complex. AI-ready data, meaning data that is accurate, consistently defined and reachable for the use in question, then has to be maintained as source systems change.
Checking output is a standing job, and more so once agents act on it. The Stanford AI Index 2026 (April 2026) reports that AI agents still fail roughly one in three attempts on structured benchmarks, and that documented AI incidents rose to 362 in 2025, from 233 in 2024. Yet in New Relic's 2026 Observability Forecast (September 2026; 2,575 IT and engineering leaders and practitioners in 24 countries, surveyed in April and May 2026), 25% of organisations had AI agents in production with no monitoring in place. The NCSC's May 2026 guidance on agentic AI, written with international partners, advises starting with tightly bounded pilots, giving agents the least access for the shortest time, monitoring their behaviour and planning for incidents.
Where does the gap between AI tools and AI results open up?
In the businesses we work with, the gap opens in four places. Data nobody owns: the same customer is defined differently in three systems, and nobody is responsible for fixing it. Systems that do not connect: the AI tool sits beside the process, so people move data between them by hand. Output nobody checks: there is no agreed standard for good enough, so errors surface late. Nobody to run it after launch: the project team moves on, and the tool drifts, breaks quietly or costs more than planned.
Survey evidence is consistent with this, though it shows association only. In McKinsey's State of AI 2026 (August 2026, 1,719 respondents), nearly three-quarters of the 6% it classes as AI high performers had fundamentally redesigned workflows around AI, against a quarter of other respondents, and one in five respondents said running costs had constrained their AI use. Scale is uneven too: 40% of respondents from organisations with revenue above $1 billion reported scaling AI agents, against 22% at smaller organisations.
UK firms also name the AI skills gap. The Office for National Statistics (July 2026) found a lack of expertise among the most cited barriers to AI adoption, reported by around 18% of businesses with 100 to 249 employees. Training existing staff is the most common response. It helps people use the tools, but in our experience specialist capabilities such as data engineering and AI operations need people whose job it is. For the failure patterns, read why most AI implementation fails.
Build, buy or blend: how should you source AI capability?
Build means hiring and developing your own team. Buy means outsourcing to an IT services provider, a technology company or a specialist AI and data firm, or hiring consultants for a project. Blend means outside specialists where they add most, and a dedicated team for the continuing work. AI is also making the build route more attractive: in McKinsey's 2026 survey, 32% of respondents said their organisation had decided against buying at least one software product or feature because it could be built in-house with agentic coding tools. Whatever you build, someone then has to test, secure, maintain and eventually retire it.
| Route | Suits | Watch for | Where the knowledge ends up |
|---|---|---|---|
| Build in the UK | Continuing work that needs to sit close to UK customers and decision-makers | We see specialist technical hires take six to twelve weeks or longer in the UK, and trained people are in demand elsewhere | With you, while the people stay and their work is documented |
| Buy from a supplier or consultants | Specialist design and build work, or a time-boxed assessment or pilot, that you do not have the skills for in-house | How changes are priced after launch, and what happens to the capability when the contract ends | With the supplier, unless handed over |
| Blend | Specialist help for design and the first build, a dedicated team for the continuing work | A clear handover and a named owner from day one | With you, if the dedicated team stays and documents its work |
Where you are on the journey changes the answer. If you have not started, an honest look at your data, systems and skills, with some outside advice, may be enough for now. If you are exploring, specialist help can shorten a pilot, but start the data foundations with people who will stay. If you are committed, the continuing work needs a dedicated team, in the UK, offshore or both.
Compare the full cost of each route over several years, including recruitment and management time. ITJobsWatch puts the median advertised UK salary for a permanent data engineer at £70,000 in the six months to 25 September 2026, with the top tenth of adverts at £100,000 or more. In our experience, salaries in Bengaluru, the largest engineering pool we recruit from, often run 40 to 70% below the UK, especially for mid-to-senior roles. Labour cost savings typically run at 30 to 50% against the UK in Cape Town and Johannesburg and 40 to 50% against Western European rates in Iași, and in São Paulo operating costs for equivalent roles typically run 30 to 50% below the UK. Build the business case on your own numbers.
What should stay owned by your business?
Whoever does the work, four things should stay yours.
- Decisions and accountability: what the AI may do, where a person must approve and who answers for the outcome. The ICO's guidance on AI and data protection says that senior management remains accountable for the data protection implications of AI and cannot delegate them to data scientists or engineering teams.
- Your data and what it means: what counts as a customer, an active contract or a margin, and which data the AI may use.
- The knowledge of how your systems behave: which datasets are reliable, why each pipeline and integration works as it does, and what breaks when a source changes. Without documentation, that knowledge leaves with the people who hold it.
- The assets: integrations, configurations, prompts, test cases and documentation. Make sure your contracts give you ownership of them and access to them.
This is why we build teams that work inside your business. Under the Embedded Offshore Team Model, the team works in your systems, to your priorities, and reports to your managers, while we handle employment, workspace and local infrastructure. Because the team works in your systems, what it learns builds up where you can reach it, provided it is documented and shared, as with any team. Our guide to dedicated development teams and staff augmentation compares where the knowledge ends up under each model.
If the work touches personal data, UK GDPR applies throughout. The same ICO guidance, under review following the Data (Use and Access) Act 2025, says that in the vast majority of cases AI involves processing likely to result in a high risk, which legally requires a data protection impact assessment. Giving an offshore team access to personal data is usually a restricted transfer, because the team's employer is a separate legal entity from you. Romania is covered by UK adequacy regulations, so no extra transfer safeguard is needed, though your other UK GDPR duties still apply; South Africa, India and Brazil need a safeguard such as the International Data Transfer Agreement plus a transfer risk assessment. Take advice on your own position.
When does an embedded offshore team make sense for AI capability?
An embedded offshore team makes sense when four things are true: the work continues after launch, it needs people who learn your data and systems over time, the skills are slow or expensive to hire in the UK, and you want the capability to stay yours. For a one-off design exercise a specialist firm is usually the better buy, and small volumes may be absorbed by training your existing team.
Much of this is work we already build teams for. When Evora Global's consultants were spending too much of their time on data work, we built a data team in Iași for its data management and operations and a software team in Bengaluru for its software roadmap; the Iași team cut weeks from the effort of getting data to Evora's clients. For Opinium, we built a Cape Town team focused on data processing and cleaning, which, over a 12-month period, returned about 20 days a month to insight consultants in London, New York and Amsterdam. For Open Energy Market, which needed to build its software before revenue caught up, we built a dedicated development team in Bengaluru.
We build these teams in five hubs in four countries: Cape Town and Johannesburg in South Africa, Iași in Romania, Bengaluru in India and São Paulo in Brazil. We can build the capabilities in this article in any of them, from data engineers, data analysts and BI developers to QA engineers, site reliability engineers and SOC analysts, and the right hub for your team depends on a discussion with you about your business. In our experience, hiring takes four to six weeks, or four to eight in São Paulo, from starting the search to the candidate accepting the offer; we see the UK equivalent take six to twelve weeks or longer. More in our guide to the best offshore location for a technology team.
A practical AI readiness checklist
Use these questions as a quick AI readiness assessment for any planned use. Each "no" is a capability gap to close before you scale. Data protection cannot wait that long: where personal data is involved, any required data protection impact assessment must be completed before the processing starts, including in a pilot.
- Data: do we know which systems hold the data this use needs, who owns each one, and who fixes the data when it is wrong?
- Integration: can the AI tool or agent reach those systems with the right permissions, and who maintains the connections?
- Testing: who checks the output, against what standard, before launch and after every change?
- Security and data protection: do we know what the AI can see and do, and where personal data is involved, have we assessed whether a data protection impact assessment is required and completed one where it is?
- Operations: who monitors accuracy and cost, handles incidents and applies updates after launch?
- Ownership: do we own the integrations, configurations and prompts, and could we change supplier without losing them?
- People: which of these can our team absorb, which must we hire, buy or augment, and how long will that take?
For each gap, decide whether to build, buy or blend, and name who will own it a year from now. Book a discovery call to discuss the capabilities you need to build, augment or scale around your AI.
Frequently asked questions
What is AI readiness?
AI readiness is whether a business has the data, systems, people and processes to make AI tools work reliably in its own operation. In practice that means six capabilities around the tools: data engineering, data quality, integration, quality assurance, security and AI operations, each with a named owner after launch.
What does AI-ready data mean?
AI-ready data is data an AI tool can use accurately and safely for a specific purpose: complete, consistently defined across systems, current, reachable with the right permissions and lawful to use. Getting there takes data engineering and data quality work, and staying there is continuing work as source systems change.
What is the difference between an AI project and an AI capability?
An AI project delivers something once, such as a pilot or an integration. An AI capability is the continuing work that keeps it useful: maintaining data pipelines, fixing data quality, testing after changes, monitoring accuracy and cost, and handling incidents. Plan and budget for both.
Should we build, buy or blend our AI capability?
Buy specialist help for one-off design and build work you cannot do in-house. Build or blend the continuing work, because data engineering, data quality, testing and AI operations need people who know your systems and stay, such as a dedicated team in the UK or offshore that owns the capability after the project.
Can an offshore team support our AI and data work?
Yes, where the work is ongoing. Your offshore team works within your systems, processes and standards, under your direction and management. Potentiam can build teams of data engineers, data quality specialists, data analysts and BI developers, QA engineers, site reliability engineers and security specialists in any of our hubs, supported by our local infrastructure. We agree the right hub with each business, typically hiring in four to six weeks (four to eight in São Paulo) from search to offer acceptance. If personal data is involved, plan the transfer safeguards first.
Sources: Skills England, What works for AI upskilling in the UK (July 2026); GFT Technologies and Wakefield Research, CIO survey (September 2026); Deloitte, agentic AI survey (August 2026); Stanford HAI, AI Index Report 2026 (April 2026); New Relic, 2026 Observability Forecast (September 2026); NCSC, agentic AI guidance (May 2026); McKinsey, The State of AI in 2026 (August 2026); ONS, Artificial intelligence in UK businesses: 2023 to 2026 (July 2026); ICO, Guidance on AI and data protection; ICO, international transfers (January 2026); ITJobsWatch, Data Engineer (six months to 25 September 2026).
