Build plus subscribe: how mid-market firms should buy AI capability

If you buy AI capability the way you buy a website build, you will regret it within a year. If you buy it the way you buy SaaS, you will regret it within a quarter.

The two default commercial shapes that mid-market firms use to buy technology were designed for different things. Neither of them fits AI capability well. The version that does fit is a hybrid, and the firms getting the most value out of AI in 2026 are the ones who have quietly figured this out.

Why the project fee breaks

The classical consulting shape is a fixed-scope, fixed-price project. Someone builds you a thing. The thing lands. The consultants leave. You own the thing.

For a website, a data migration, or a system integration with a clean end state, this shape works. The thing you bought is the thing you will still have in three years, with minor maintenance.

For AI capability, this is not what happens. The model landscape moves. The prompt that worked in March is a bit stale by August. A capability that was expensive last year is cheap this year, and cheap this year is free next year. Vendors deprecate APIs. Client workflows evolve. A "finished" AI system that gets no further attention decays measurably within six months.

The project shape produces a moment of value and then a slope of decay. In a category that compounds if it stays live and only compounds if it stays live, that is exactly the wrong shape.

Worse, the project fee anchors both parties in the wrong direction. The consultant is incentivised to over-scope the build (because that is where the margin sits) and under-invest in the transition (because that is where the effort sits). The client ends up with a bigger thing than they needed and less of the operational literacy they needed to keep it going.

Why pure SaaS also fails at this size

At the other end of the shelf, there is the SaaS pitch. "Buy our AI agent platform. Configure it yourself. Renew annually."

There are two problems with this at mid-market scale.

The first is that the interesting AI use cases are not generic. The bottleneck you are trying to unblock has your firm's specific systems, your firm's specific process, your firm's specific data. Off-the-shelf platforms handle the generic 40% of that surface and leave the specific 60%, which is where the value lives, to your team to figure out. If you had the team to figure it out, you would not have bought the platform.

The second is that SaaS AI is not portable. You commit to a vendor's data model, their integration story, their fine-tune choices, their pricing curve. When one of those changes in a way that hurts you (and one of them will), your options are pay more, accept less, or migrate. Migration off an AI platform is much worse than migration off a CRM. The prompts, the tool definitions, the audit trails, and often the model behaviour do not transfer.

SaaS AI works at the small end, where the specific 60% does not matter, and at the enterprise end, where a client can absorb the operational cost. In the middle, it is a poor fit.

The hybrid that works

The shape that fits AI capability at mid-market scale is a fixed-scope build for the initial capability, then a monthly subscription that keeps that capability alive and improving, with any genuinely new work quoted separately.

Concretely:

A fixed-scope initial build. Two to six weeks. One specific bottleneck, well defined, with an agreed outcome. Priced up front, paid on delivery. This produces the first working capability the client will operate.

A monthly subscription. Covers cloud hosting, platform operations (patching, monitoring, security), incremental improvement (template tweaks, small config changes, minor UX adjustments, prompt updates as the model landscape moves), a regular steering call, and async access for questions between calls. This is not big money. Anchor tier engagements sit around $2,000 (AUD) per month. Core tier engagements land in the $60,000 to $120,000 per year range.

Anything genuinely new is quoted separately. A new agent, a new integration, a new job type, a new user group. Fixed-scope, fixed-price, delivered inside the subscription cadence and then absorbed into the ongoing service. The client always sees the number before committing.

The boundary between the subscription and the extension work is the load-bearing detail. Everything already built stays alive and keeps drifting toward better on the subscription. Anything new is a scoped conversation. Both parties know which side of the line any given ask sits on.

What clients get from this shape

The economics stay honest for both sides.

The consultant is incentivised to keep the capability alive, because the recurring revenue only holds if the client keeps deriving value. The consultant is not incentivised to over-scope the initial build, because the profit is not concentrated there.

The client gets predictable operating cost, ongoing access to the technical depth that built the capability, and a partnership that survives the natural drift of the AI landscape. When Claude 5 replaces Claude 4, the subscription keeps the models current without a new contract. When a template changes, it is a config update, not a change request.

The client also gets to grow the capability at their own pace. A firm that scoped a modest first build in month one can decide in month six that it wants a second agent, a third integration, or a new job type. The extension is quoted, delivered, and folded in. The overall shape stays coherent.

What consultants have to give up

This shape is not free for the consultant either.

It requires being disciplined about scope and honest about extensions. The temptation to quietly do new work under the subscription is real, and it kills margin. The discipline is to always price new work explicitly, even small pieces.

It requires being present. A subscription-based partner has to show up every month, listen, and act. If the steering call becomes ceremonial, the client will feel it inside two quarters.

It requires giving up the large upfront fee that traditional consulting shops rely on for cashflow. The subscription is smaller in year one than a big project would have been. The compounding value is in year two, year three, year five.

For a solo or boutique practice, this is the right trade. For a large consulting shop with billable-hour targets, it is heresy.

The sustainability point

The reason this shape matters is not that it feels more modern. It is that the compounding value of AI capability only shows up if the capability stays live. A capability that is left alone after the initial build produces diminishing returns and eventually stops working.

A capability that is operated, improved, and evolved on a predictable cadence produces the opposite. Small improvements accumulate. New use cases open up. The infrastructure that was built for one bottleneck ends up serving the next one at a fraction of the cost. The subscription pays for itself several times over inside the first year, and by year two it is the operating platform the firm quietly relies on.

The mid-market firms that will be having the good AI outcomes in 2028 are the ones structuring their engagements this way now. The ones that bought a big project last year and are looking at a stale system this year will be doing another project soon.

Or, if they are smart, they will be flipping the shape.

---

This piece completes a three-part series that started with where AI agents are actually earning their keep and continued with the deterministic-pipeline rule. For the full canonical guide, see AI Agentic Development.