Most AI projects do not fail on the technology.
They fail on decisions made before the build. The work here is choosing what is worth building, confirming your data can support it, comparing vendors on the same terms, and knowing the cost before everyone is using it.
of AI projects fail to deliver the business value they were funded for, about twice the failure rate of comparable IT projects
RAND Corporation, 2024
of organizations running generative AI pilots report no measurable return to the P&L
MIT Project NANDA, 2025
of companies abandoned most of their AI initiatives in 2025, up from 17 percent the year before
S&P Global Market Intelligence, 2025
The pattern behind those numbers is consistent. No shared definition of success, data that was never prepared for the use case, integration nobody scoped, and an invoice that arrives at a size no one modeled. None of that is a model problem, and all of it is decided before the first pilot runs.
The forcing event is almost always a date, not a technology.
The call usually comes after a board meeting or a budget cycle. Someone senior has committed to an AI plan, and the commitment arrived before the scope did. By then three departments have bought their own tools on expense reports, and at least one pilot is sitting stalled because security asked a question that was never answered.
The evaluation is not a sales path for one platform. Everything is written down, including the options that were rejected and the reason. The result is a document you can hand to finance and legal without translating it first.
Four decisions the work covers
- 01Which use cases are worth building
- 02Whether your data can support them
- 03How vendors compare when tested on the same material
- 04What the licensing costs at full adoption
There is no product to sell. Compensation comes from the vendors through the TSD model, so the engagement carries no cost to you and the recommendation follows the requirements rather than a quota.
Start where it applies
Leadership has asked for an AI plan and nobody agrees on scope
Security or legal stopped a pilot and it was never resolved
Every demo looks impressive and none of them are comparable
Pricing is per seat and nobody can model it at full adoption
Use Case Selection
An AI plan is a list of decisions, not a direction. With scope open, every department nominates its own priority and none of them can be compared. This produces one ranked list with a named owner on each line.
How candidates get ranked
- Every AI tool already in use, including the ones on expense reports
- Monthly volume of the tickets, documents, or calls each use case touches
- Whether the output has to be right every time or mostly right
- The person who signs off before an AI answer reaches a customer
The use cases that survive are almost never the ones named in the meeting. They are the ones where a person is already reading the same document twice a day and can tell you exactly what a wrong answer costs.
Data Readiness Review
Pilots rarely stop on model quality. They stop when someone asks where the data goes and no one has the answer in writing. This review produces that answer before the next pilot starts.
What gets traced
- Where each data set physically sits: your tenant, the vendor's, or a subprocessor's
- Retention terms written in the contract, not the claim on the website
- Which permissions the tool inherits, and whether they match the user's own
- Written confirmation that your prompts and files do not train the model
Most stalled pilots are not blocked on policy. They are blocked because the permission model of the source system was never built to be read at machine speed, and a search tool will surface every file a user technically has rights to.
Vendor and Model Evaluation
A demo is the vendor's data, the vendor's prompt, and the vendor's best day. Comparison requires your material and the same test run against each tool. This builds that test and scores what comes back.
What gets compared
- The same fifty of your own documents run through every tool
- What each model does when it does not know the answer
- Which integrations exist today and which are on a roadmap slide
- Whether you can move to another model later without rebuilding the work
Give every vendor the same ten questions you already know the answers to, and make three of them unanswerable. The tool that says it does not know is worth more than the one that is confidently wrong, and only one of those failure modes ever appears in a scripted demo.
Licensing and Cost Review
Per seat pricing is quoted at pilot size and paid at company size. The distance between those two numbers is where the budget breaks. This models the cost at each stage of rollout before you sign.
What gets modeled
- Cost at pilot, at one department, and at every employee
- Message or token caps, and the overage rate once a team passes them
- What the price becomes at renewal when the year one discount ends
- Features you already pay for inside the Microsoft or Google agreement
Adoption is the risk in both directions. Low usage makes a per seat contract indefensible at renewal, and heavy usage on a consumption meter produces an invoice nobody budgeted, so the agreement needs a floor and a ceiling before it needs a discount. Consumption pricing is where this shows up now, because token and message metering turns adoption into a variable cost, and the invoice arrives after the usage rather than before it.
How the engagement runs
Understand your environment
Current inventory, contract dates, sites, and what the business actually needs the technology to do.
Define where you want to go
Target state, growth requirements, and success criteria, all agreed before a single vendor is contacted.
Evaluate solutions
The requirement goes to every provider that can meet it, and the responses come back on terms that can be read side by side.
Define a solution and negotiate terms
Pricing, term length, service levels, and exit language are settled before signature rather than discovered on the first invoice.
Oversee activation
We stay in the project through installation, porting, and acceptance, and we escalate on the client's behalf when dates slip.
What you walk away with
A ranked use case list with the owner, the data source, and the volume for each
A data path memo showing where information travels for every tool under review, with the gaps named
A vendor scorecard built on your own documents, with the same criteria applied to each product
A licensing model showing cost at pilot, at department rollout, at full adoption, and at renewal
Questions we get
- We already have Microsoft 365 Copilot licenses. Does that settle the vendor question?
- It settles part of it. Copilot is strong where the work already lives in Outlook, Teams, and SharePoint, and it inherits your existing permissions, which is either the advantage or the exact problem depending on how your file shares were built. It does not reach the use cases that sit inside your line of business systems, and it bills per seat whether or not the seat uses it. The evaluation is usually about what you add around it, not whether you keep it.
- Security and legal stopped our last pilot. What actually gets it moving again?
- In most cases the blocker was never a policy disagreement. Legal was asked to approve a demo and had no document describing where the data goes, how long it is held, and who else touches it. Get those answers from the vendor in writing, along with the subprocessor list and the training terms, and the review becomes an ordinary contract review. That takes days instead of the months a stalled pilot usually sits.
- Departments already bought their own tools. What happens to those?
- You count them before you touch them. Most of what people bought is doing real work, and canceling it without a replacement costs you the goodwill you need for the rollout. The ones that matter are the tools where company data was uploaded into an account no one owns, and those get moved onto a contract with your name on it. The rest can usually run until renewal.
- How are you paid if the engagement costs us nothing?
- Compensation comes from the provider through the TSD model, which stands for technology services distributor. When a contract is signed, the provider pays a distribution fee that would otherwise go to its own direct sales team. Your pricing comes off the same rate card either way.
- What if the right answer is staying where we are?
- That gets recommended when it is correct. A renegotiated contract with the incumbent, or leaving a working system alone, counts as a finished engagement and gets the same work as a migration.
- When should we start relative to our contract end date?
- Roughly six months out. Auto-renewal notice windows commonly close 60 to 90 days before term end, and starting inside that window removes most of the leverage. Check your specific renewal clause, since the window is the constraint rather than the end date.
The first conversation is a scoping call, not a pitch.
Bring your current Microsoft or Google license counts, a list of the AI tools departments have already bought, and the one use case leadership keeps naming. That is enough to say whether there is anything worth pursuing.
The first conversation is a scoping call, not a pitch.
We ask what you have, what is expiring, and what is not working, then tell you whether there is work here worth doing. If Elk Run is not the right fit for the decision in front of you, we will say so.