Skip to content

Guide · 9 min read · Sep 18, 2026

In-house, agency or partner: how to staff an AI project

An AI project needs a business owner and a process owner from inside your company, plus AI engineering, integration, data, cloud and change skills that you can hire, contract or bring in through a partner. For a first workflow, a specialist partner working with named internal owners is usually the fastest route, and hiring makes sense once there is steady work to keep a team busy. Whichever route you take, your company should own the code, the accounts and the documentation.

Tuaha JawaidFounder, business lead

Key takeaways

  • Two roles must come from inside: the business owner accountable for the result and the process owner who knows the work.
  • Hiring builds lasting capability but is slow and hard to get right before you know how much AI work you will have.
  • Freelancers and general agencies can build well, but you may carry the coordination and the AI-specific testing.
  • A specialist partner suits the first one or two workflows, and a hybrid model grows internal ownership release by release.
  • Own the code, the cloud and model provider accounts and the documentation from day one, whoever does the building.

Once a company has chosen an AI workflow worth building, the next question is who will build it, and who will keep it running a year later. The options range from hiring a team to handing everything to an outside firm, and each one fails in a predictable way. This guide sets out the roles involved, compares the staffing models and shows how to keep the knowledge inside your business whichever route you take.

What roles does an AI implementation need?

Think in roles, not headcount. On a first project one experienced person may cover two or three of these, and some are needed for only a few weeks.

RoleWhat they doWho usually fills it
Business ownerAccountable for the outcome, the budget and trade-off decisionsAlways internal: a department head or executive
Process ownerKnows the workflow and its exceptions, approves the design, signs off testingAlways internal: the manager who runs the work
AI/ML engineerChooses models, designs prompts and retrieval, builds evaluation tests and guardrailsHire or partner
Integration engineerConnects the ERP, CRM, email and document systems through APIs, with error handlingInternal IT, hire or partner
Data engineerBuilds data pipelines, fixes quality problems, sets up access to the right sourcesInternal data team or partner
Cloud and security engineerDeploys on AWS or Azure, manages identity and access, logging and cost controlsInternal IT, partner or specialist
Change and training leadInvolves users early, runs training, tracks whether people actually use the solutionInternal operations or HR, with partner support

Two terms in that table are worth defining. Retrieval means fetching the relevant company documents so a model answers from them rather than from general knowledge. Evaluation tests are a fixed set of real examples with known correct answers, used to check the system before release and after every change.

The last row is easy to skip and expensive to miss. BCG's 2024 survey of 1,000 senior executives found that around 70 percent of AI implementation challenges stem from people and process issues, 20 percent from technology and only 10 percent from the algorithms. The business owner, the process owner and the change lead are where most of that 70 percent is won or lost, and none of those roles can be fully outsourced.

What are your options for staffing an AI project?

Hire in-house

Your own team knows the business and stays with the system. It is also the slowest route, and the hardest to get right before you know how much AI work you will have. Two reference points from the US Bureau of Labor Statistics: the median annual wage for software developers was $135,980 in May 2025, and for data scientists it was $120,230, with data scientist employment projected to grow 35 percent from 2025 to 2035. Medians cover whole occupations, not the specialists an AI project needs most.

Salary is not the full cost either. In private industry, benefits made up 30.0 percent of employer compensation costs in June 2026. Add recruiting time, management attention and the risk that a single key hire leaves.

Freelancers

Individual contractors can be excellent for a defined piece of work, such as one integration or one data pipeline. The catch is coordination: with three freelancers, someone in your company has to design the whole system, review the work and make the pieces fit. Continuity is fragile, because the person who understands the code may be booked elsewhere when something breaks.

General development agency

A software agency brings a managed team and a delivery process, which removes much of the coordination burden. Many are strong at web and mobile applications. AI workflows add different disciplines, such as measuring output quality, handling uncertain results and designing human approval steps, so ask to see how the agency tests AI behavior rather than assuming app experience transfers.

Large consultancy

Large firms offer breadth, established governance and the capacity to run multi-country programs. For a company of 50 to 500 people they can be heavy, and the engagement model often favors strategy phases over working software. Check who will actually do the work. A large consultancy makes most sense when AI is one part of a wider transformation program.

Specialist AI implementation partner

A smaller firm focused on AI and automation typically covers the engineering, integration and cloud roles in one team and works directly with your business and process owners. The strengths are focus and speed on the first workflows. The risks are dependence on a small outside team and, if the handover is poor, a system your staff cannot change. Our guide to choosing an AI implementation partner covers how to vet one.

Hybrid

Most companies end up here: internal owners for the business roles, an outside team for the specialist build and a planned transfer of day-to-day ownership. A hybrid only works if that transfer is designed in from the start.

How do the staffing options compare?

The table compares the models for a company running its first one to three AI workflows. Cost profiles are described qualitatively because rates vary widely by market and scope.

OptionSpeed to startCost profileAccountabilityKnowledge retentionMain risk
Hire in-houseSlow, recruiting comes firstFixed ongoing payroll, whether or not there is workClear and internalHigh, while people stayHiring before you know what you need
FreelancersFast for a defined taskVariable, pay for time usedSplit across individuals, coordinated by youLow unless documentedGaps between pieces and availability
General development agencyModerateProject fees for a managed teamWith the agency, for the agreed scopeMedium, depends on handoverApp skills without AI testing practice
Large consultancyModerate to slowHighest overhead, often phasedWith the firm, often split across phasesMedium, often held in documentsStrategy that outruns delivery
Specialist implementation partnerFast for the first workflowsProject fees, plus optional supportWith the partner, against agreed measuresMedium to high with a planned handoverDependence on a small outside team
HybridFast, then gradualProject fees shifting toward payrollShared, with named internal ownersHigh if the transfer is plannedUnclear lines of ownership

When should you hire, and when should you partner?

Hire when AI work is continuous rather than occasional. If you can list more work than one team could deliver in the next year, and the systems involved are central to how you compete, permanent staff will repay the time it takes to recruit them. RAND's 2024 study of why AI projects fail advises leaders to be prepared to commit each product team to solving a specific problem for at least a year. That makes a useful test: if you cannot see a year of work, a permanent AI team will struggle to stay busy.

Hiring also needs someone to hire against. Without a technical leader who can write the role, judge candidates and review their work, a first AI hire often ends up isolated.

Partner when:

  • You are building your first one or two workflows and still learning which AI work pays off.
  • There is no internal technical leader yet to direct an AI hire.
  • You need scarce skills for a short period, such as a cloud deployment, a security review or a complex ERP integration.
  • A business deadline will not wait for a recruiting cycle.

Do neither when the job is standard. If a mature product already handles the workflow well, buying it is usually faster than staffing a build. Our guide to build, buy or integrate covers that decision.

How do you stop knowledge walking out the door?

People will leave whichever model you choose: employees change jobs, freelancers take other contracts and partners finish engagements. The protection is to make ownership and understanding part of the deliverables rather than an afterthought.

  1. Own the accounts. Cloud accounts on AWS or Azure, model provider accounts with OpenAI or Anthropic, code repositories and domains should be in your company's name, with an employee holding administrator access. If a partner hosts the system, the contract should say who owns what and what happens at the end.
  2. Own the code in writing. Under US copyright law, software written by an outside contractor is not automatically yours. The US Copyright Office explains that a commissioned work counts as a work made for hire only if it falls into one of nine listed categories and both parties sign a written agreement saying so. Software is not named among those categories, so ask your lawyer for an explicit assignment of rights in any development contract. Rules differ outside the US.
  3. Make documentation a deliverable. Architecture and data flow diagrams, configuration and prompts, integration details, a runbook for common failures and a list of known limitations. Name it in the contract with an acceptance step.
  4. Keep the evaluation tests. They are how your team will know whether a future change made the system better or worse.
  5. Pair from the start. Assign an internal technical contact who joins design decisions, reviews changes and has the same access as the builders.
  6. Hand over against criteria. Before the outside team steps back, your team should ship a real change and handle a practice incident while help is still available.
  7. Close access cleanly. At the end, remove external user accounts and rotate credentials and API keys.

What does a phased staffing approach look like?

A phased model lets a partner move quickly on the first workflow while internal ownership grows with each release. Treat the phases below as a pattern to adapt, not a fixed timeline.

PhasePartner roleInternal roleMove on when
1. First workflowDesigns, builds, tests and deploysBusiness and process owners decide and test; a technical contact shadowsThe workflow is live, baseline measures are tracked and documentation is accepted
2. Second workflowBuilds alongside internal staff and reviews their workThe technical contact makes changes; hiring needs are defined from real workInternal staff have shipped a change to production
3. Internal ownershipAvailable for specialist work or under a separate support agreementRuns operations, monitoring and routine improvementsThe internal team handles changes and incidents unaided

The second phase lets you hire against evidence of which skills the work actually needs, rather than a guess made before the first build.

Illustrative example: Consider a regional distributor with 180 employees whose customer service team re-keys purchase orders from emailed PDFs into its ERP. With no AI experience in house, it brings in an implementation partner to build order extraction, with a person approving every order that fails a validation check. The IT manager joins every design review and owns the cloud and model provider accounts from the first day.

For the second workflow, routing delivery exceptions, the IT manager makes the integration changes while the partner reviews them. Only then does the company write a job description, and it is for an integration engineer rather than the data scientist it had first assumed it needed.

How Kastling approaches staffing on an AI project

Kastling is a founder-led team: a business lead and an AI engineering lead, with specialists, including AWS and Azure engineers, brought in when the work needs them. We act as the specialist partner in the hybrid model above, working with your business and process owners rather than replacing them.

Engagements start with a free discovery call. For AI and operations work, a separately scoped paid audit commonly follows, although it is not required for every project. The proposal sets out scope, milestones, measures of success and who owns the accounts, access and operating costs. Deployment can use your cloud account or ours, and the arrangement is written down either way.

During development we test against real business scenarios with the people who will use the solution. Integration and training include agreed documentation and a defined transition into day-to-day ownership. Maintenance is optional and covered by a separate agreement. See AI Integration & Automation for what we build, or start with an AI readiness assessment to decide what to staff for first.

Questions

Can one person cover all the technical roles on a first AI project?

Sometimes, on a narrow workflow with one or two systems and no sensitive data. It gets risky when the same person designs the AI behavior, builds the integrations, secures the cloud setup and supports users, because nobody checks their decisions and nobody covers when they are away. At a minimum, have someone else review security and access choices.

Should our IT team build AI automation themselves?

IT teams are well placed to own integrations, identity and access, and day-to-day operations. The newer skills are designing and testing AI behavior: judging output quality, handling uncertain results and deciding where a person must approve. A common split is for IT to own the platform and accounts while a partner or specialist hire covers those skills until the team has learned them.

Who should own the API keys and model provider accounts?

Your company, with billing in your name and administrator access held by an employee. Outside builders can work with scoped keys that you can revoke at any time. That keeps usage costs visible and means ending a contract never takes a live system offline.

Is it cheaper to hire than to use a partner?

It can be over several years of steady work, because a team spreads its cost across many projects. For a first workflow it often is not, once recruiting time, benefits, management and the risk of a poor hire are counted. Compare the cost of a year of delivered, working systems rather than day rates against salaries.

How long should an outside partner stay involved?

Until your team can make a change and handle an incident without help. That depends on the complexity of the system more than the calendar, so set it as an exit criterion rather than a date. Some companies then keep a light support agreement for specialist work or busy periods.

Sources

  1. US Bureau of Labor Statistics: Occupational Outlook Handbook, Software Developers, Quality Assurance Analysts, and Testers
  2. US Bureau of Labor Statistics: Occupational Outlook Handbook, Data Scientists
  3. US Bureau of Labor Statistics: Employer Costs for Employee Compensation, June 2026
  4. BCG: AI Adoption in 2024, 74% of Companies Struggle to Achieve and Scale Value
  5. RAND: The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed (2024)
  6. US Copyright Office: Circular 30, Works Made for Hire

Guide · 14 min read

AI workflow automation: a practical guide for operations teamsAI workflow automation puts AI models inside a defined business process to read, classify and prepare work, while fixed rules, integrations and named people handle the steps that must be predictable. It suits repetitive, document-heavy or request-heavy work such as intake, routing, document checks and approvals. Start with one measurable workflow, keep consequential approvals with an accountable person, and expand only once the numbers show it works.Read

Article · 9 min read

How to calculate the ROI of AI automationCalculate the ROI of AI automation by measuring the current process first, then valuing hours saved at loaded cost, errors avoided and faster cycle times, and setting that against the full cost to build, run and support the system. Express the result as a payback period under conservative assumptions, and count saved hours as money only when they change hiring, overtime or contractor spend.Read

Article · 9 min read

AI document processing for invoices, contracts and formsAI document processing combines three techniques: OCR to turn images into text, layout models to understand where values sit on a page, and language models to extract and interpret fields that vary between documents. The technique matters less than what surrounds it: validation rules that check extracted values against your own records, confidence thresholds that decide what a person sees, and a review queue somebody owns. Measure field level accuracy and the share of documents that pass without a human touch, not a single headline accuracy number.Read

One useful email a month.

New guides, tools and practical notes on putting AI to work. No sales sequences.

Bring us a workflow.

Tell us where the work slows down. We will help you see where to start.

Book a discovery call