Applying AI to corporate applications: assess, prepare and plan before the easy button

The pressure, and the easy button

The board asks. The CEO asks. A competitor announces something. Within a quarter the CIO is expected to have an AI strategy, a budget line and a rollout date. Under that pressure the path of least resistance is the platform the organization already pays for: turn on the AI features in the productivity suite, most often Microsoft Copilot, because it is the easy button. It requires no architecture conversation and it produces something to show leadership in weeks.

That decision makes sense for some clients, and we say so when it does. For many others it is a decision made for the sake of speed, and the questions it skips come back later as cost, exposure or a pilot that nobody can measure.

The four questions that get skipped

1. Is the data optimized for AI consumption?

AI is only as useful as what it can reach and understand. Most corporate data lives in file shares, SharePoint sites, ticketing systems, ERPs and spreadsheets that were never organized for a model to consume. Duplicates, stale versions, broken permissions and undocumented fields all surface the moment an assistant starts answering from them. Data readiness is the unglamorous work that determines whether the answers are right.

2. Where and how does the data feed the model?

Connectors, retrieval pipelines, vector indexes, agents with tool access: each is a path by which corporate data reaches a large language model (LLM). Those paths have to be designed, not discovered. Which systems are in scope, what is excluded, how permissions are inherited, and what the model is allowed to retain are architecture decisions that the easy button makes for you by default.

3. On-premises, or a cloud frontier model?

Some workloads belong with a frontier model in the cloud: broad knowledge, fast iteration, the best general capability. Others belong on infrastructure the organization controls, because of regulation, contract, data sovereignty, latency or steady-state cost. Most organizations need a position on both, and a way to route the right work to each, before the first use case is chosen.

4. How will the data be protected?

Once AI is in the environment, data moves in new ways: into prompts, into indexes, out through agents and plug-ins, and across SaaS boundaries. Protecting it means intelligent security platforms that watch for data exfiltration and misuse in those new paths, not just a policy document. This is the question most often left entirely unasked, because it slows the rollout down.

A perspective we share: outcomes start with the foundation

We like working with CBTS on this, and we think they have the message right. Their position is that real AI outcomes start with the right foundation, and that most AI programs fail on data readiness, infrastructure that cannot scale or a strategy nobody funded, rather than on the technology. Their Forge AI offering brings advisory, data readiness, engineering and managed operations together so an organization can deploy practical AI with predictable costs and a return it can actually measure.

That matches what we see. The organizations getting results did not pick the smartest model. They did the foundation work first, and their pilots were designed around a number that already existed.

Assess, prepare, plan: what happens before AI is introduced

Our role is to work with the client and our partners to properly assess, prepare and plan before AI enters the environment.

1. Assess. Where the data lives and what condition it is in; which applications and processes already have measurable baselines; what the security stack can and cannot see; what the organization already pays for that includes AI features.
2. Prepare. The data readiness work, the permission clean-up, the decision on on-premises versus cloud models for each class of workload, and the security controls that will monitor the new data paths.
3. Plan. A first use case with a written baseline and a target; the architecture for how data feeds the model; the pilot group; the cost at scale, not just at pilot size; and who owns the outcome.

Then the platform decision, which may still be the easy button, now chosen rather than defaulted to.

Where the measurable results come from

With the foundation in place, the use cases that pay back first are the ones already instrumented: contact center agent assist and call summaries, service desk triage, meeting summaries pushed into the task system, invoice and contract extraction with a review queue, policy search that answers with citations. For each, write one sentence before the pilot: “Today X takes Y; after the pilot we expect Z, measured by this report.” If the sentence cannot be written, choose a different use case. That sentence, and the foundation under it, is the difference between an AI strategy and an AI subscription.