Start with the simplest route that can complete the agreed work. Better use of ChatGPT or Claude may be sufficient for a task someone starts and reviews. Configured projects or assistants may help when instructions and reference material repeat. A custom workflow becomes worth considering when a specific requirement remains unmet, such as a controlled handoff between systems or a purpose-built review interface.
The product name alone will not settle the decision. Describe the inputs, accepted output, people, permissions and next step, then test which route supports them with a reasonable operating burden.
This guide separates three choices that are often bundled into “we need AI”. It offers a practical route worksheet rather than a ranking of products.
First distinguish the tool from the complete job
Suppose someone uses AI to prepare a customer reply. The complete job might also require finding the current order, checking a policy, recognising an exception, obtaining approval and recording what was sent.
A useful draft solves part of that work. Whether you need more depends on the effort and consequences of the remaining steps. The person may already handle them effectively. Or the business may need a consistent handoff that a chat alone does not provide in its current configuration.
Write the requirement without prescribing the technology: “Prepare a response from the permitted order and policy records, show missing information, and leave the customer commitment to the coordinator.” That gives each proposed route the same job to prove.
Route one: improve how people use the tool
This is worth testing when a person can supply the permitted context, inspect the answer and carry it to the next step without creating significant extra work or ambiguity.
Useful improvements might include a clearer task instruction, a good example, a review checklist or practice recognising unsupported statements. The test is whether these changes improve the actual task. Buying a different subscription does not itself establish that result.
For example, a small team might need help turning an approved outline into an internal meeting brief. If the outline is complete and a person reviews the brief, better use of a current tool may be enough. The task does not automatically require a new application.
For product selection, use ChatGPT vs Claude for business. For consistent team use, see the rollout guide.
Route two: configure repeatable context and instructions
When people keep supplying the same instructions or looking for the same approved material, a configured working space may be useful.
OpenAI describes ChatGPT Projects as a way to organise related chats, files and instructions. Its GPT guidance distinguishes instructions from uploaded knowledge and makes creation and sharing subject to account and workspace permissions. Anthropic's Projects guidance describes project knowledge and instructions, with organisational sharing subject to the applicable setup. Check current availability in the actual account; these are different product features, not interchangeable entitlements.
Configuration still needs an owner. Someone must decide which source is authoritative, replace stale reference material, review sharing and test what happens when the source does not answer the question.
A project containing documents is not automatically a dependable company knowledge system. Test retrieval, source references, access and missing information. The company-knowledge guide provides a source register and test cases.
Route three: commission a connected workflow or custom build
Consider a build when you can state a useful requirement that the simpler routes cannot meet adequately in the selected environment. That could be a specific review screen, reconciliation across permitted sources, persistent case status or a controlled update to another business system.
Before commissioning it, ask for a demonstration of the gap. Can an existing supported integration meet it? Could the process change? A custom build may be justified, but its maintenance, access management and failure handling become part of the decision.
Define where AI contributes and where ordinary rules are sufficient. A deterministic check can validate a required field. A person can decide an exception. AI might prepare material between those steps. The complete system should make each responsibility inspectable.
Compare routes using the same worksheet
Complete one column for each realistic option. Write the evidence or the next test, rather than assigning a confident score to an untested capability.
If columns don't fit, scroll sideways. Keyboard: focus the table and use the arrow keys.
| Requirement | Current tool and better practice | Configured project or assistant | Connected workflow or custom build |
|---|---|---|---|
| Accepted output | Can the operator produce and review it? | Does reused context improve it? | Which unmet requirement would the build address? |
| Source information | What must someone supply? | Who maintains reference material? | How is permitted retrieval controlled and checked? |
| Actions | What does the person do next? | Which tools are actually enabled? | Which actions need separate approval? |
| Exceptions | Can the operator detect and resolve them? | How are missing or conflicting sources exposed? | Where do failed cases stop and return? |
| Operating owner | Who maintains the task method? | Who maintains instructions and access? | Who maintains code, connections and service checks? |
| Total cost | Seats, practice and review effort | Configuration, seats and upkeep | Build, provider usage and ongoing operation |
| Evidence needed | A representative task test | A reuse and permissions test | A complete workflow and recovery test |
The worksheet may show that two routes should coexist. A team can use a general tool for flexible preparation and a narrow connected system for one repeated job. That does not require rebuilding every activity inside the custom system.
A fictional example: where the boundary changes
This example is invented and contains no client results. Imagine a service team preparing an internal job brief from a completed intake form.
At first, an operator supplies the form, asks for a brief and checks it. A shared instruction and review checklist may solve the immediate need. If repeated context is the main burden, the team can test a configured project using approved reference material.
Later, the business wants an accepted brief to create a job record with attachments and a visible exception queue. The team would first inspect its existing software. If a necessary handoff remains unsupported, it can scope a connected build around that gap.
The decision changes because the job changes. It does not follow from a belief that custom AI is always the more advanced or valuable choice.
Keep subscriptions, API access and implementation separate
As checked on 2 November 2026, OpenAI maintains separate billing systems for ChatGPT and its API platform. Eligible ChatGPT subscriptions may offer API credits, but eligibility, activation and the receiving API organisation must be checked; a ChatGPT seat does not grant API organisation access.
Anthropic similarly describes Claude and Console accounts as independent, while documenting API-credit benefits for certain plans. Check current entitlements and conditions instead of assuming either that all API usage is included or that no subscription can provide credits.
Professional implementation work is another cost. Your quote should identify the capability, provider costs, account responsibilities, tests, training and ongoing operation. A credit benefit does not pay for every part of building and maintaining a system. Use the cost guide to compare the complete scope.
Choose a route you can test and operate
Run representative cases through the proposed route, including missing and contradictory information. Compare the complete preparation, review and correction effort. Ask the intended operator to demonstrate what they would do when the system cannot complete the job.
Build n Bloom's ChatGPT + Claude Implementation service helps make this choice and implement the agreed work. Your business's provider accounts, subscriptions and access are handled under the applicable contracts and agreed scope; a services engagement is not a claim that we resell a platform or supply its entitlements.
A free 25-minute Fit Call can begin with one task. You receive a short reviewed provisional Fit Note, with timing agreed during the call. The next step may be a direct Build or Connected Build, an optional Blueprint or better use of what you already have.