Choose the assistant that handles your actual work with acceptable data terms, manageable permissions, and a review process your team can follow. At Applied Margin, we would shortlist Copilot for a Microsoft-based office, Gemini for a Google-based office, and ChatGPT or Claude when a separate assistant fits the task. That is a starting list, not a ranking. Test the exact product and plan before putting customer information into it.
Start with one team, a small set of approved examples, and one default assistant. Compare the complete job: preparing inputs, getting a draft, checking it, and returning the result to the working record. Keep customer-facing work and consequential decisions under human review. A tool that produces a polished answer but creates more checking or awkward copying can lose to a less impressive demonstration that fits the day.
What should you compare before choosing an assistant?
Imagine an HVAC office preparing a customer update from a technician's notes. The useful question is whether the coordinator can produce an accurate, approved message without losing the job reference, inventing a completion time, or exposing information unnecessarily. A general conversation about which model sounds smarter will not establish that.
Use the following table to compare candidates. Fill it with evidence from your own test and the vendor's current documentation. An unanswered data or access question is a reason to pause that use case, even when the writing is good.
| Decision | What to establish | What a useful test looks like |
|---|---|---|
| Existing work | Where the source material and final record live | Complete one task from the normal starting point, including the handoff back |
| Data terms | Exact account, plan, approved data, agreements, retention, and access | An administrator records the settings and the applicable vendor terms |
| Everyday output | Whether the result meets your actual standard | A staff member checks facts, omissions, tone, and required formatting |
| Connections | Which files and systems the assistant can read or change | Verify a user cannot obtain material outside their approved permissions |
| Administration | Who manages membership, permissions, and departures | Demonstrate how access is removed when an employee leaves |
| Operating effort | Total work required to produce a checked result | Include preparation, corrections, copying, and recovery from a bad answer |
Do not turn the table into a contest with a winner at any cost. If none of the tools can meet the data requirements or produce a result people can verify, simplify the task or leave it with the existing process. There is no requirement to standardize on an assistant before you have a useful job for one.
Does your current office suite decide the answer?
It helps decide what to test first. A team that already manages work in Microsoft tools should examine the relevant Copilot offering. A team working in Google Workspace should examine the relevant Gemini offering. Existing identities and familiar work surfaces can make a trial easier to organize, but they do not establish that a specific feature, license, or integration meets your needs.
Ask your administrator to show the actual workflow in your own environment. Which product is enabled? Which account is signed in? What sources can this person reach? Does the assistant perform the needed task under the terms you reviewed? A personal product with a familiar brand name is not interchangeable with your organization's approved service.
ChatGPT and Claude belong on the shortlist when a separate workspace is a practical way to draft, examine documents, or prepare internal work. Compare them using the same approved examples and acceptance criteria. Avoid declaring one universally better for writing or analysis: your material, instructions, verification needs, and operators determine whether the result is useful.
For a small agency, the deciding factor might be whether staff can prepare a client update from an approved report without inventing an explanation for a change. For a multi-location service business, it might be whether the assistant handles location-specific rules without combining them. These are illustrative selection questions, not claims about results from clients.
Which data promises need a closer look?
Separate model training, storage, access, and connected services. They answer different questions. OpenAI states that its business offerings do not use business data for model training by default; its documentation separately describes retention and access. Read the section for the product you are evaluating, rather than applying an API or Enterprise statement to every ChatGPT account. OpenAI enterprise privacy.
Anthropic also describes a no-training default for commercial products, with exceptions when customers choose to share data, including certain feedback. Its retention documentation distinguishes stored conversations, API use, contractual arrangements, and model-specific requirements. A no-training promise does not mean that the provider never stores or processes the material. Anthropic commercial training policy, commercial retention policy.
Microsoft's enterprise data protection documentation says covered Copilot interactions are not used to train foundation models. It also distinguishes web-search queries and agents from the main service's contractual protections. Verify the precise offering and enabled features with your administrator. Microsoft enterprise data protection.
Google distinguishes Gemini provided as a Workspace core service from Gemini provided as an additional service; the applicable data protections differ. A work email address alone does not answer which terms apply. Verify your edition and service status. Google's Workspace account comparison.
For every candidate, record whether searches, extensions, connectors, and external actions introduce another processor or different terms. Verify the contract, enabled features, and administrator settings together. Treat a security badge or a business subscription as evidence to investigate, not approval for every document in your office.
Keep this review proportionate to the proposed work. Public brochure text and a patient record do not need the same authorization. The customer-data policy field note provides a way to classify examples before a trial. Begin with public or genuinely synthetic material while unresolved questions are checked.
How do you run a useful comparison without making it a project?
Choose a short trial period and name the decision-maker before testing. Use a handful of representative tasks, including an incomplete input and a case the tool should decline to finish. This is a suggested evaluation method, not a promise that a fixed number of examples establishes reliability.
- Write the finish line. Define what an acceptable result contains, what it must never claim, and who approves it.
- Prepare permitted inputs. Use public, synthetic, or specifically approved information. Give candidates equivalent context rather than comparing one informed answer with one bare prompt.
- Run the whole task. Have the person who normally does the work operate the tool, including the return to the CRM, document, or inbox.
- Record the corrections. Note missing facts, unsupported claims, permission problems, and the work required to fix them. Keep the original source beside the output.
- Choose or stop. Select a default only if it passes the practical and data checks. If it fails, identify whether the task, instructions, source material, or tool needs changing.
Measure review effort as well as generation speed. An answer that needs five minutes of checking instead of ten may be useful. An answer that needs a specialist to reconstruct every claim may not be. Treat any time estimate as a trial observation, not a projection of profit across the company.
The broader point is already covered in the work around the model: the selection test should include the work surrounding the answer. You do not need a permanent model leaderboard to choose an assistant for one operating task.
What should you standardize after the trial?
Standardize the approved workspace, the instructions, the reference material, the reviewer, and the place where finished work belongs. Write a short task guide that another person can follow. Include an example of an acceptable result and an example that must be stopped or corrected.
Shared context should describe the business without becoming a collection of private customer records. The company.md and customer.md approach can help organize company rules and general customer expectations. Give those files an owner and a revision date. Approve their contents before uploading them; a useful filename does not make sensitive information safe to share.
Keep task instructions and evaluation examples in a business-owned location outside a single person's chat history. That helps with continuity and makes later switching less disruptive. Do not assume exports preserve every tool-specific setting, connection, permission, or automation. Document those dependencies separately.
Apply the Operator Standard: the person running the task must understand it, use the workflow, verify the result, override it, and recover when it goes wrong. This is where buying seats and building a working habit separate. People need practice and an owner they can ask when the input falls outside the written example.
When is using two assistants reasonable?
Two tools can make sense when the second has a defined job the first does not perform adequately under your requirements. For example, one approved workspace might support routine office drafting while another supports a separate team with different materials and review needs. Both still require appropriate terms, access, and ownership.
Write down the reason, who may use each tool, what data may move between them, and when the exception will be reviewed. Do not let staff copy a confidential conversation into another service simply to see whether its answer sounds better. Each transfer is another data-handling decision.
Review your default on a schedule you can maintain, and sooner after a material change in vendor terms, permissions, connected services, or the workflow. Switching because a new release is interesting is different from switching because the current setup cannot do the job. Preserve what works while testing a replacement with permitted examples.
What is a sensible first engagement?
If the task and controls are clear, your own team may be able to run the comparison. When people need guided practice, the AI Workflow Sprint is $1,500 fixed for one team of up to five people. It includes a short leadership kickoff, two working sessions, two to three practical workflows, reusable instructions, a prioritized opportunity map, and one follow-up session.
Session lengths and scheduling are agreed before booking. Custom software, integrations, and subscriptions are separate, and the Sprint has no automatic credit toward a later build. A deeper paid assessment remains optional; a defined implementation can go straight to scoping. Let’s talk about your business and the task you want the team to handle more reliably.
FAQ
Is the free version of an AI assistant OK for work?
It can be suitable for permitted public or synthetic material, but a free or personal account should not become an unapproved route for customer data. Check the exact service, account settings, and company policy. A paid personal subscription also does not establish organizational approval. Use the data requirements to decide what belongs in the tool before comparing the quality of its answers.
Should we wait for the next model before choosing?
Wait if the current tools cannot meet a requirement that matters to your workflow. Otherwise, test the tools available to you on a bounded, useful task. Keep instructions, source material, and evaluation examples outside a single chat so you can compare a later release. A release announcement alone is not evidence that switching will improve your team's checked work.
Can we switch from ChatGPT to Claude, Copilot, or Gemini later?
Usually you can reconsider the choice, but migration effort depends on what you have built around it. Preserve business-owned instructions and reference files, and document permissions, integrations, retention obligations, and where completed work is stored. Test the replacement before moving customer information. Do not assume saved chats, custom configurations, or connected workflows transfer unchanged between vendors.
Does a small business need an enterprise plan?
Headcount alone does not answer that question. List the access, retention, contractual, support, and regulatory requirements for the intended use. Then verify which exact offering satisfies them and what configuration or agreement is required. A smaller organization may have demanding data obligations; another may only need an approved workflow using public material. The plan label is not the acceptance test.
Which assistant writes the best customer messages?
Choose by testing the messages your team actually needs, using permitted source material and the same acceptance criteria. Check factual accuracy, tone, missing qualifications, and how long a person takes to approve the draft. No assistant should independently decide your promises, prices, or advice. Customer-facing messages remain human-reviewed before sending, whichever tool produces the first version.
Put it to work
Install the steps, prompts, human approval rules, and scoreboard in Get Found in ChatGPT.