What AI really costs, what it returns, and how to pick the partner who delivers it

What the platform and the engagements cost, in public.

We publish prices because unpublished pricing is a tax on your evaluation time. A/DVNT Enterprise™ — the licence that runs the deployed operating layer — ships in two commercial modes under AMD-107. Mode 2, A/DVNT-Managed, is $150K per year: we operate the governance and operating layer on your behalf. Mode 3, Customer-Operated, is $295K per year: your team operates it and licenses the IP.

The service ladder is priced separately because it is a different purchase. A/DVNT Advisory™ starts at $18K and is the diagnostic entry point. Assess (4 weeks), Govern (6 weeks), Architect (6 weeks), and Enable (12 weeks) start from $28K each. You do not need all five, and most organisations start with Advisory and one other.

Below the flagship, A/DVNT Business is the mid-market platform edition on the same substrate, sized down. A/DVNT Cloud has a free tier for entry tools, which exists so a team can prove value without opening a procurement cycle.

Read the ladder as a ladder. Advisory tells you what to buy; the engagements build the capability; the Enterprise licence runs it. Buying the licence without the capability is how organisations end up paying for a runtime nobody operates.

Enterprise Mode 2

$150K/yr — A/DVNT-managed operating layer, Axxon bundled.

Enterprise Mode 3

$295K/yr — customer-operated, IP licensed to your team.

Advisory

From $18K — the diagnostic that scopes everything else.

Assess · Govern · Architect · Enable

From $28K each — 4 to 12 weeks depending on the rung.

The five line items missing from almost every proposal.

Licence and implementation are the visible costs and rarely the decisive ones. The lines that move a business case are the ones that appear in month four, after the budget has been signed against a number that only covered the first two.

Inference and usage is the obvious one, and the shape matters more than the rate: token costs scale with adoption, so a successful pilot produces a bill that looks like a failure of forecasting. Model your cost per completed workflow, not per token, and re-forecast at each tier of adoption.

Data preparation is the largest single surprise. Access, classification, retention rules, and the cleanup nobody sequenced. If the pilot needs a source that has never been governed, the governance work is a project in its own right and belongs in the plan rather than in the overrun.

The remaining three are organisational. Evaluation and monitoring is a standing cost, not a launch cost. Change management — the training, the workflow redesign, the period of parallel running — is routinely underfunded by an order of magnitude. And integration debt accrues quietly whenever a point tool is bought without a substrate: each one adds an identity model and an audit gap you will pay to reconcile later.

Inference at scale

Costs rise with success. Forecast per workflow, per tier, not per token.

Data preparation

Access, classification, retention. The most common source of a doubled timeline.

Evaluation & monitoring

A standing operational cost for the life of the workflow.

Change management

Training, redesign, parallel running. Chronically underfunded.

Integration debt

Every ungoverned point tool is a reconciliation bill deferred to year two.

There is no universal AI ROI number — returns track your tier.

Any vendor quoting a single ROI multiple is quoting someone else's baseline. Return depends almost entirely on the adoption tier you start from, because the constraint is different at each one and the money is unlocked by removing the constraint, not by adding the model.

At the lowest tier the return is time recovered on high-frequency manual work, and it shows up in weeks. It is real, it is modest, and it is the evidence that funds the next stage. At the middle tiers the return shifts to cycle time — how long a case, claim, ticket, or deal takes end to end — which is where the numbers become material to a P&L.

At the upper tiers the return is capacity: work the organisation could not previously take on at all, and risk cost avoided because governance is enforced rather than audited after the fact. That last component is real money in regulated sectors and is systematically excluded from vendor ROI models because it is hard to attribute.

Build the case in three lines and defend each separately: recovered time, valued honestly at loaded cost and discounted for the portion that does not convert to output; cycle-time improvement, tied to a specific throughput or revenue metric; and avoided risk cost, expressed as a reduction in exposure with the reasoning shown. A board will accept a smaller number that is defensible over a larger one that is not.

The AI Audit benchmarks your posture against the 2026 YC + a16z dataset and projects payback by tier, which gives the model a baseline that came from somewhere other than the vendor proposing the work.

Entry points that do not become a rebuild.

Not every organisation should start at the flagship, and starting small is only a mistake when the small thing is architecturally disposable. The question to ask of any entry-level option is what happens at the upgrade: do you migrate, or do you keep the substrate and change the shape?

A/DVNT Cloud is the free tier for entry tools — enough to run a real workflow and produce real evidence without a procurement cycle. A/DVNT Business is the mid-market platform edition: the same governance substrate as Enterprise, sized to a smaller estate and a smaller operating team.

Because both run the same operating layer, moving up is a licence and scope change rather than a re-implementation. Policy, audit trail, and identity model carry forward. That is the specific property to insist on from any vendor at this price point, and the one most cannot offer.

For services, Advisory at $18K is deliberately the smallest useful commitment: it produces a scoped plan you can execute yourself, take to another partner, or hand back to us. A partner unwilling to sell you a diagnostic without the delivery attached is telling you something about the diagnostic.

Filter for the substrate they leave behind.

Most AI consulting engagements end with a deck and a Slack channel that goes quiet. The evaluation framework that avoids this is short, and every question in it is about what still exists ninety days after the team leaves.

Ask to see the operating layer they will leave behind. If it is a bespoke build with a bus factor of one, you are buying a deck with an implementation phase. Named, versioned, licensed IP is a materially different purchase from a custom repository that only its authors understand.

Ask where governance is enforced. If the answer is a policy document and a review board, governance ends at the boundary of everyone's attention. Ask what audit trail you keep and whether you can query it without them — a partner who owns your evidence owns your renewal.

Ask whether your team can operate it without them, and take the answer seriously. Genuine enablement is a specific deliverable with a curriculum, an intake path, and named internal operators — A/DVNT Enable™ is a twelve-week engagement precisely because it is a real body of work, not a knowledge-transfer call at the end of a build.

Finally, ask for published pricing and a scoped diagnostic. Both are cheap signals of the same thing: a partner confident enough in the work to let you evaluate it in pieces rather than committing to a programme on the strength of a logo slide.

The substrate test

What runs in production after they leave, and who else can maintain it?

The enforcement test

Where is policy evaluated — at dispatch, or in a review meeting?

The evidence test

Can you query your own audit trail without the vendor?

The exit test

Can your team operate it, and what is the named plan to get them there?