top of page

Claude or Copilot? You're Asking the Wrong Question

  • Jun 21
  • 6 min read

What actually determines whether AI agents deliver value for your business — and why the vendor you standardise on matters less than the questions nobody is asking.


Claude or Copilot is no longer a strategic decision
Claude or Copilot is no longer a strategic decision

It usually comes up in a leadership offsite. Someone has read another vendor comparison, watched another product keynote, or sat through another sales pitch — and the question lands on the table: "Are we a Claude organisation or a Copilot organisation?" Whether the answer is Claude or Copilot rarely turns out to be the question that matters.


A working group forms. Spreadsheets get built comparing pricing tiers, model benchmarks, integration depth with existing Microsoft tenancy, and vendor roadmaps. Weeks go into the decision. A vendor is selected. A rollout plan is drafted. The business announces, internally, that it has chosen its AI platform.


Six months later, the agents are running, the licences are paid for, and the results are mixed at best. Some teams have found genuinely useful workflows. Others have quietly stopped using the tool because it could not see the data it needed, or because nobody could explain what it was allowed to access in the first place. The vendor decision consumed months of leadership attention. The questions that actually determined whether the rollout worked were never asked.


This is not a hypothetical. The pattern is consistent enough to name: organisations are treating "which AI vendor" as the strategic decision, when the vendors themselves have already started resolving that question — and the decisions that remain genuinely strategic are being skipped entirely.



Why "Claude or Copilot" Feels Urgent (and Why It's Already Out of Date)


The Claude-versus-Copilot framing made sense eighteen months ago, when the two ecosystems were genuinely separate and a business had to pick a lane. That is no longer the world we are operating in.


Microsoft's Frontier program now puts Claude directly into mainline Copilot chat alongside OpenAI's models, with Copilot automatically applying the model it judges best for a given task rather than the end user choosing. At the data platform layer, Databricks' Unity AI Gateway applies the same logic to AI workloads run on its platform — routing requests to the most appropriate model based on task complexity, quality requirements, and cost, with budget controls applied across that routing. Both of the major platforms a mid-sized business is likely to be standardised on are converging on the same idea: model selection is becoming infrastructure, decided dynamically per task, not a once-off platform commitment made by a steering committee.


This matters because it means the multi-month "which vendor" decision many businesses are currently running is solving a problem the market is already solving for them. It is not that vendor choice is irrelevant — contractual terms, data residency, and existing tenancy still matter operationally. It is that vendor choice has stopped being the lever that determines whether AI agents actually create value. Something else is the lever, and most leadership conversations have not moved on to it yet.



The Five Questions That Actually Determine Whether AI Agents Work


1. Governance Reach — Who Decided What This Agent Can See?

Every agent operates with some set of permissions, whether or not anyone deliberately designed them. The question that matters is whether those permissions were a considered decision or an inherited default. An agent that quietly has the same data access as the staff member who configured it is not a feature — it is an unreviewed risk sitting inside a workflow that looks, from the outside, like it is working fine.


This is a common and easy-to-miss failure mode: a finance-team agent originally scoped to expense reports inherits read access to an entire shared drive when an upstream folder structure changes. Nobody reviews the permission after the change. The agent keeps functioning normally — and remains quietly over-permissioned — until someone happens to check, often well after the change that caused it.


2. Routing and Model Flexibility — Has Your Vendor Already Decided This for You?

If your AI platform increasingly routes between models automatically — as both major ecosystems now do — locking your organisation's strategy to a single named model is solving for a constraint that is disappearing. The more useful question is whether you have visibility into which model handled a given task and why, particularly for anything that touches customer data, financial reporting, or compliance-relevant decisions. Routing without observability is a black box with a vendor logo on it.


3. Cost Visibility — Do You Know What Each Agent Is Costing You, and Why?

At Databricks' Data + AI Summit in June 2026, the expanded Unity AI Gateway release focused heavily on cost: unified spend visibility across model providers, granular cost attribution by user and team, and hard spend caps that automatically stop requests when a budget is exceeded. Microsoft's own Frontier rollout has introduced budget alerts and per-user spend thresholds alongside its multi-model routing. When two of the largest platform vendors in this space are both building dedicated cost-control infrastructure rather than just new model capabilities, it is a reasonable signal about where the operational risk actually sits.


Token and compute costs for agentic workflows scale differently from traditional software licensing — a handful of poorly scoped agents running frequent, large-context tasks can produce a cost profile nobody budgeted for, discovered only when the invoice arrives. Budget controls and per-agent cost attribution need to exist before agents go into production, not after the first unexpected bill.


4. Data Readiness — Is There Anything Reliable Underneath the Agent?

This is the question we wrote about in detail in our piece on tokenmaxxing: flooding an AI system with raw, ungoverned data and trusting it to make sense of the mess. The same foundational problem applies here in a different shape. No model — Claude, Copilot, or whatever replaces them next year — can produce trustworthy output from a data layer that is inconsistent, ungoverned, or undocumented. Vendor selection cannot fix a foundation problem. It can only determine how confidently the system produces wrong answers on top of one.


5. Workforce Augmentation Path — Is There an Actual Plan, or Just a Rollout?

Organisations that get genuine value from AI agents tend to treat adoption as a maturity path: agents augment specific, well-scoped tasks first, prove themselves, and only then take on more autonomous responsibility, with governance scaling alongside that autonomy at every step. Organisations that get more limited value often gave every team a licence and a deadline, with no equivalent plan for what "more capable" should mean or when an agent has earned more trust. A platform rollout is not the same thing as an adoption strategy, and the gap between the two is where a meaningful amount of wasted spend tends to live.



What the Path Forward Looks Like

None of this means vendor choice is meaningless, or that the Claude-versus-Copilot conversation should never happen. It means that conversation belongs several steps later than where most organisations are currently having it — after governance reach is defined, after cost visibility exists, and after the data layer underneath the agents has been assessed honestly.


Organisations that pick the vendor first, figure out governance and cost later — typically end up retrofitting controls onto a system that is already in production and already creating exposure, which is a harder and more expensive position to work from than building the controls first.



What To Do Next

If your organisation is currently in the middle of a vendor evaluation, the most useful thing you can do before that decision is finalised is an honest look at the five questions above against your current environment — not your intended environment, the one you actually have today.


Most of this work — mapping current data access, identifying governance gaps, and getting visibility into cost exposure — does not require picking a vendor first. It holds regardless of which platform you eventually standardise on.



Cypher Agency is a boutique data and integration engineering firm helping mid-sized businesses build reliable, governed data and integration environments — without the cost of building an internal team.



References

Comments

Commenting on this post isn't available anymore. Contact the site owner for more info.
bottom of page