Every mid-market company exploring AI faces the same fork in the road. Three options, each with a persuasive pitch.

The AI consultancy says: "We'll figure out what to build, then build it. You get senior people from day one." The dev shop says: "Tell us what you need, and we'll ship it fast at a competitive rate." The internal hire pitch goes: "Build your own team. Own the IP. Control the roadmap."

Each of these sounds reasonable in a conference room. And each of them is correct, in certain circumstances. The problem is that most companies pick their path based on the wrong signal. Sometimes it's whoever happened to pitch last. Sometimes it's whoever quoted the lowest day rate. Sometimes it's the CTO who insists on owning everything in-house because that's what worked at their previous company, which had 10x the engineering budget.

I spent years as Chief Risk Officer at Navi, where every technology investment went through a financial stress test before we committed capital. When I look at the AI consultancy vs dev shop vs in-house question, I see a capital allocation decision that most companies treat like a procurement exercise.

Procurement asks: "Who gives us the best rate?" Capital allocation asks: "Where does this dollar generate the highest return, adjusted for risk and time?"

These are fundamentally different questions, and they lead to fundamentally different answers.

Gartner projects worldwide AI spending will reach $2.5 trillion in 2026 alone, with AI services representing one of the fastest-growing segments. Yet according to RAND Corporation research, roughly 80% of AI projects fail to reach production. That failure rate doesn't discriminate by engagement model. Companies working with consultancies, dev shops, and internal teams all fail at alarming rates.

The difference is in the failure mode. What breaks, how much it costs when it breaks, and how quickly you can recover. Those variables change dramatically based on which door you walk through.

So before we compare rates and timelines, let's agree on the actual question: given your company's current AI maturity, project pipeline, and internal capabilities, which model gives you the best shot at production AI that generates measurable business value within 18 months?

Let's be honest about what each option delivers, including the parts that don't make it into the sales deck.

AI Consultancy

A good AI consultancy brings strategy and execution under one roof. You get senior practitioners who have seen dozens of AI implementations across industries. They'll challenge your assumptions about what to build, help you sequence projects by business impact, and then actually build the thing.

The tradeoff: higher day rates, typically $2,000-$4,000 per day for senior consultants. Smaller teams, which means scope has to be focused. You're paying for judgment and speed, and the best consultancies will tell you when an AI project isn't worth pursuing at all.

The risk profile: you're dependent on the consultancy's talent. If their A-team gets pulled to another client, quality drops. You also don't build internal capability unless the engagement is explicitly designed for knowledge transfer.

Dev Shop

Development agencies and outsourced AI teams optimize for build velocity. Give them a clear spec and they'll deliver working software quickly, often at 40-60% of consultancy rates. The best dev shops have strong engineering processes, solid QA, and reliable deployment pipelines.

The tradeoff: you need to provide the strategic direction. Someone on your team has to define what gets built, manage the backlog, and make architectural decisions. The dev shop executes your vision. If your vision is wrong, they'll execute that too, and you'll get a polished product that solves the wrong problem.

The risk profile: scope creep is the silent killer. Without strong internal management, dev shop projects drift. Deloitte's enterprise AI survey found that while 83% of executives are leveraging AI as part of outsourced services, tangible benefits have been limited due to challenges in governance and contracting, with unclear requirements driving the majority of cost overruns.

In-House Team

Building an internal AI team gives you maximum control. Your people, your IP, your roadmap. Over a multi-year horizon with sufficient project volume, this is almost always the lowest cost-per-project option.

The tradeoff: the upfront investment is significant and the ramp time is brutal. Recruiting a capable AI/ML engineer takes 4-6 months in the current market. Once hired, they need another 3-6 months to understand your data, your domain, and your infrastructure. You're looking at 6-12 months before your first production output.

The risk profile: you're betting that AI project volume will justify the fixed cost. If the first project underwhelms (and many do), you have expensive talent sitting idle while leadership debates whether to fund project two. Attrition is the other killer. AI talent turns over at roughly 20-25% annually according to LinkedIn's workforce data, and losing your lead ML engineer nine months in can set you back to zero.

Numbers clarify things that narratives obscure. Let's build an 18-month total cost of ownership for each model, using a representative first AI project: a customer churn prediction model integrated into an existing CRM workflow.

The table below summarizes the numbers, but the story behind them matters more than the totals.

With a consultancy, the range depends on data complexity and integration requirements. The higher end includes significant data engineering work. Hidden costs are minimal: $10K-$20K in internal team time for stakeholder alignment, data access provisioning, and UAT. The consultancy handles the rest.

Dev shops look cheaper on paper, but hidden costs close the gap. Someone on your team needs to spend 15-20 hours per week managing the engagement — a $40K-$60K shadow cost over the project lifecycle if you value their time honestly. And if you haven't validated that churn prediction is the right first project, or if your data isn't ready, you'll burn 4-8 weeks of dev shop time on false starts. Surprisingly close to the consultancy number once you include these hidden costs.

The in-house model carries a cost that nobody accounts for: opportunity cost of time. Your first production model won't ship for 6-12 months. In an 18-month window, you might get one project to production. The consultancy or dev shop would have delivered in months 2-5, giving you 12+ months of value from the deployed model.

If that churn model saves $50K per month in retained revenue (a conservative estimate for a mid-market company), the consultancy option generates $600K in value during the same 18 months where the in-house team generates $300K or less. McKinsey's 2024 State of AI report found that companies using external partners for their initial AI projects reached positive ROI 2.4x faster than those who built exclusively in-house. Speed to value compounds.

Cost analysis tells part of the story. The rest depends on where your company sits today and where it needs to be in two years.

The consultancy wins when:

You don't yet know what to build. If your AI strategy is a slide deck with five potential use cases and no clear priority, you need someone who can evaluate each option against your data readiness, technical infrastructure, and business impact. A consultancy compresses months of internal debate into weeks of structured analysis.

You need strategy and execution to move together. The handoff between "the strategy firm" and "the implementation team" is where most AI projects die. When the same team that identifies the opportunity also builds the solution, the feedback loop between business logic and technical architecture stays tight.

You want production AI in under 90 days. Consultancies with pre-built frameworks and experienced teams can go from kickoff to deployed model in 8-12 weeks. Try that timeline with a new hire who starts in month three.

The dev shop wins when:

You know exactly what to build. If your internal team has validated the use case, prepared the data, defined the requirements, and designed the architecture, a dev shop is the most cost-effective way to get it built. You're buying hands, and good dev shops have plenty of capable ones.

Your internal team can manage the vendor. This means a technical product manager or engineering lead who can review code, evaluate model performance, and make scope decisions weekly. Without this person, the dev shop engagement becomes a game of telephone.

The project is well-scoped with clear boundaries. Building a recommendation engine with a defined input schema, known data sources, and agreed-upon success metrics? Perfect dev shop project. "Use AI to improve our customer experience" with no further specification? That's a consultancy conversation.

The in-house team wins when:

AI is becoming core to your product. If you're embedding ML models into your customer-facing product and plan to iterate on them continuously, in-house capability becomes a competitive advantage that you can't outsource forever.

You have ongoing volume. Three or more AI projects per year changes the math entirely. The fixed cost of an internal team gets amortized across multiple projects, and the team gets faster with each one because they know your data and domain intimately.

You can absorb the ramp time. If your competitive position allows for 6-12 months before your first AI capability reaches production, and your leadership team has the patience to fund a team through that learning curve, in-house will be the most cost-effective option over a 3-5 year horizon. Forrester's 2025 analysis of AI operating models found that companies with mature in-house AI teams (3+ years) reported per-project costs 60% lower than those relying on external partners.

After watching dozens of mid-market companies navigate this decision, a common pattern has emerged. The companies that get the best results don't pick one door and commit forever. They sequence their approach.

Phase 1: Consultancy for the first 1-2 projects (months 0-6)

Bring in a consultancy to identify the highest-impact AI opportunity, build and deploy it, and prove value to the organization. This phase accomplishes three things simultaneously: it generates business results, it educates your internal team on what good AI development looks like, and it gives leadership confidence to invest further.

The consultancy should be explicitly tasked with knowledge transfer. Every architecture decision, every data pipeline choice, every model evaluation metric should be documented and walked through with your internal stakeholders. You're buying capability, and you should insist on keeping the knowledge.

Phase 2: Selective in-house hiring (months 4-9)

Once you have a production AI project generating measurable value, you have the evidence to justify internal hires. Start with one senior ML engineer who can own the existing model and scope the next project. This person will cost less to recruit because you can show them a real, deployed AI system rather than pitching a vague AI vision.

Phase 3: Dev shops for commodity work (ongoing)

As your AI portfolio grows, certain work becomes routine: data pipeline maintenance, integration builds, dashboard development. This is perfect dev shop territory. Well-defined scope, clear requirements, and no need for deep AI expertise. Your internal team defines the spec and reviews the output.

Phase 4: Consultancy for new frontiers (as needed)

When you're ready to tackle a new category of AI problem (moving from predictive models to generative AI, or from internal tools to customer-facing products), bring the consultancy back for the strategic and technical heavy lifting. Then repeat the cycle.

This sequenced approach typically costs 20-30% more than going all-in on a single model in year one. Over three years, it costs 30-40% less because you avoid the most expensive failure modes: building the wrong thing (consultancy prevents this), building it slowly (dev shop prevents this), and failing to build internal capability (phased hiring prevents this).

I've sat in enough budget meetings to know that AI investment decisions often get reduced to "how much?" and "how long?" Those are the wrong first questions. Here are the five questions that separate good capital allocation from expensive experiments.

1. What's the payback period for each option?

Map the expected business value (revenue increase, cost reduction, risk mitigation) against the total cost and timeline for each engagement model. A consultancy might cost more per month but deliver a 6-month payback. An in-house team might cost less per month but take 18 months to reach payback. The monthly burn rate is a distraction. Cumulative ROI at the 18-month mark is what matters.

2. What's the switching cost if this doesn't work?

If the consultancy engagement fails, you're out the project fee and you start over with lessons learned. If your in-house hire doesn't work out, you've lost 6-9 months of salary, recruiting costs, and the opportunity cost of delayed projects. If the dev shop delivers something that doesn't meet business requirements, you're facing a rebuild. Quantify the downside scenario for each option before you compare the upside.

3. What are we learning vs what are we renting?

Every engagement model teaches your organization something about AI. The question is what and how much. A consultancy engagement with strong knowledge transfer builds lasting internal understanding. A dev shop engagement that ships code without context builds dependency. An in-house team builds deep expertise but only in the specific problems they tackle. Score each option on capability building, because that's the compounding asset.

4. How does each option affect our operating leverage?

Fixed costs (in-house team) create operating leverage: as AI project volume grows, the marginal cost of each project decreases. Variable costs (consultancy, dev shop) scale linearly with volume. If you expect 5+ AI projects in the next two years, the leverage math favors in-house. If you expect 1-2 projects, variable cost models preserve flexibility.

5. What's the opportunity cost of the slowest option?

This is the question that changes the most minds. If your competitor deploys AI-driven pricing optimization six months before you do, what does that cost in market share? If your customer support AI goes live in Q2 instead of Q4, what's the value of those two additional quarters? Time-to-production has a dollar value, and it's usually larger than the difference in engagement fees between your options.

Analysis without action is just expensive procrastination. Here's a practical framework for moving forward based on where you are today.

If you have zero AI in production:

Start with a consultancy diagnostic. This is a 2-4 week engagement, typically $5K-$15K, where a senior AI practitioner evaluates your data readiness, identifies the highest-impact use cases, and recommends an implementation path. You'll get a concrete project plan with realistic timelines and costs, which is a much better basis for the build-or-buy decision than abstract comparisons.

This diagnostic pays for itself even if you decide to go in-house afterward, because it compresses months of internal exploration into weeks and gives your future hire a validated starting point.

If you have a successful pilot:

You've proven that AI can generate value in your organization. Now the question is how to scale. Evaluate three paths in parallel: extending with your current partner (if you used one), bringing in a dev shop to productionize the pilot, or hiring an internal lead to own the scaling effort. The right answer depends on your project pipeline. One follow-up project? Extend with an external partner. Three or more projects identified? Start the internal hire and use external partners to bridge the gap.

If you have an internal team that's struggling:

This is more common than anyone admits. You hired smart people, gave them an ambitious charter, and twelve months later the team has built impressive prototypes that can't quite make it to production. A consultancy can help here by pairing with your internal team for a focused engagement: get one project across the finish line, establish production-grade practices, and build momentum. Your team learns by shipping alongside experienced practitioners.

The one choice that's always wrong:

Doing nothing because you can't decide between the three options. Every quarter of inaction is a quarter where competitors are deploying, learning, and compounding their AI advantage. The cost comparison between consultancy, dev shop, and in-house pales in comparison to the cost of standing still.

Pick the model that matches your current maturity. Move forward. Adjust as you learn. The companies that win with AI aren't the ones who chose the theoretically optimal engagement model on day one. They're the ones who started, measured, and iterated.