The AI project you ask for and the AI project that pays off are usually two different things.
Most requests arrive shaped by cost. You want delivery to move faster, spending to drop, or agents running somewhere visible. Those are fair goals, and they still describe a tool before they describe a problem. A recent survey of nearly 6,000 senior executives found nine in ten reporting no impact on employment or productivity from AI over the past 3 years. Most of their firms were already using it.
That gap is where good AI strategy consulting earns its fee. We think the first job of any AI advisory work is to test the request against the business problem behind it. Sometimes the request survives intact. More often it gets smaller, moves to a different team, or turns into a data project first. You end up buying the AI that helps, which is the only kind worth funding.
What most AI requests are really asking for
Most AI requests ask for lower cost or faster output, and they name a technology before they name a result. You'll hear "we need agents" or "make our software delivery faster" long before anyone states which customer, process or number should improve. The request is a symptom worth reading carefully.
Picture a wealth operations lead who asks for an agent to handle client onboarding. The ask sounds specific. Underneath it sit 3 separate problems. Identity checks run slowly, staff key the same data into two legacy systems, and a compliance review queue holds every file for days. An agent will touch all 3 and fix none of them unless someone pulls them apart first.
We don't treat requests that start with cost as wrong. Cost pressure is real, and your board is asking the same questions you are. The trouble starts when the savings target becomes the project scope. A target tells you how much to save. It says nothing about where the waste lives or which fix will remove it.
Why the request drifts away from the business problem
Requests drift because new technology gets adopted as a checklist item before it gets matched to a goal. Leaders feel pressure to show AI activity, so the question becomes "how are we using AI" and stops being "what are we trying to accomplish". Activity is easy to count. Outcomes take longer to prove.
We've watched this pattern before with cloud migration tallies and mobile app launches. A bank reports how many workloads moved or how many apps shipped, and the count starts to stand in for value. AI now sits in that same seat. A dashboard showing 12 pilots looks like progress to a steering committee, even when none of them has reached production.
"A target tells you how much to save. It says nothing about where the waste lives or which fix will remove it."
| What you ask for | What the problem usually needs |
|---|---|
| You want agents running in operations this quarter. | Someone has to map the workflow first and find the steps that actually cause delay. |
| You want AI to cut the cost of software delivery. | Requirements and testing often hold the waste, so coding speed alone moves very little. |
| You want a chatbot to reduce call centre volume. | Customers usually call about a broken process, and fixing that process removes the calls. |
| You want a model to speed up credit decisions. | Clean, traceable data and a clear approval policy have to exist before a model adds value. |
| You want AI licences for every employee. | People need coaching on their own tasks, or the licences sit unused after the first month. |
Treat those rows as prompts for a conversation. The first version of any request carries assumptions nobody has checked yet. Naming them early costs you an afternoon, and it spares you a quarter of rework.
How to test an AI request against the real problem
You test an AI request by stating the business imperative first, then asking if AI is the right tool to reach it. That takes a few direct questions, answered by the people who own the process. If the answers hold up, the request is sound. If they wobble, you've saved a budget cycle.
These 5 questions do most of the work:
- What business result must improve, and who owns that number?
- What happens today, step by step, when the process runs?
- Where does the time or money go inside those steps?
- Would a simpler fix, such as a rule or a cleaner data feed, solve it?
- How will we know within 90 days that it worked?
Run that list against a claims team asking for a generative AI summarizer. The third question often shows that adjusters lose most of their time chasing missing documents, and reading files takes a small share. A summarizer speeds up the smaller part. An intake checklist with automated document requests fixes the larger one, and it's far easier to govern.
Signs that AI is the wrong tool for the job
AI is the wrong tool when the problem is unclear, the data isn't there, or a fixed rule would do the work. Those 3 signs show up early if you look for them. They're also the signs teams skip past when a deadline or a vendor demo sets the pace.
The cost of skipping them is well documented. Estimates cited in research on AI project outcomes put the failure rate at more than 80%, twice the rate of technology projects that don't involve AI. The same research lists a misunderstood or poorly communicated problem first among the root causes. Missing data and a focus on the newest technology follow close behind.
A payments team wanting a model to approve refunds under a set dollar value shows the pattern clearly. The policy already says yes below the threshold. A rule handles it with full auditability, and no model risk review is needed. Saving AI for the ambiguous cases above the line gives you a smaller build and an easier conversation with compliance. It also frees budget and risk appetite for work where judgment, language or pattern recognition matter.
What AI strategy consulting should deliver before any build starts
Good AI strategy consulting delivers a tested problem statement, a ranked set of use cases, and the guardrails each one needs before any build starts. You should leave with fewer projects than you walked in with. Each one should carry an owner, a measure and a governance path.
That's how we scope work at Electric Mind. One engagement for a large North American bank began with a 4-week AI-first current state assessment, and the mobile banking platform build came after it. The assessment came first so that a lean team could build against known constraints.
Advisory work that skips this step produces a slide deck and a pilot. The pilot runs in a sandbox, impresses a few people and stalls at the first risk review. Foundations such as principles, data access rules and AI governance sound like delay. They're what lets a regulated firm move quickly later, because nobody has to stop and argue about them halfway through a build.
Where the budget belongs once the use case is clear
Once the use case is clear, most of the budget belongs with the people who'll use the tool, and a smaller share belongs with the tool itself. Licences are the entry ticket. Adoption is the work. A model nobody trusts or understands will sit unused, however good it is.
Consider a finance department handed an AI assistant and a one-hour webinar. Usage spikes for a week and fades. Now give the same team hands-on coaching built around month-end close, with their own reconciliations as the practice material. Analysts start shaping the tool to their work, and they flag the control gaps before audit does.
We'd rather see you fund fewer use cases and train deeply than spread licences thin across every team. People who've used AI on their own tasks ask better questions of the next request. They can tell a good fit from a fashionable one. That judgment compounds, and it protects you from the next round of hype.
"Licences are the entry ticket. Adoption is the work."
Choosing the right AI project takes patience before speed
Choosing the right AI project means going slow at the start so you can go fast afterward. You state what you're trying to accomplish, check that AI fits, then commit fully. The patience is front-loaded. Everything after it moves quicker because the argument about purpose is already settled.
We've sat in enough kickoff meetings to know how hard that pause is. Budgets have dates on them, and someone senior has already promised agents by next quarter. Asking "what are we solving" at that moment can feel like obstruction. It's the most useful question in the room.
Electric Mind starts every engagement by testing the request against the problem behind it, and we stay accountable through to production. Sometimes the answer confirms what you asked for. Sometimes it points to a quick win, and sometimes to a foundation you'll build on for 3 to 5 years. Either way, you fund work you can defend. That's what the right AI advisory relationship should leave you with.
