How to Use the Copilot Researcher Agent to Compare Suppliers

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Use the Copilot Researcher Agent to Compare Suppliers.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Use the Copilot Researcher Agent to Compare Suppliers.

To compare suppliers with the Copilot Researcher agent, open Researcher under Agents in the Microsoft Copilot app, keep its Web search toggle on, and paste a brief naming the suppliers, the emails and files holding their quotes, and the criteria to score. Answer its clarifying questions, then open every citation before you decide.

The Complete Excel Course with AI Integration

From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.

By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner

Buy now

Researcher needs a Microsoft 365 Copilot licence on a work account (or Microsoft 365 Premium or Pro on a personal one), allows 25 queries per user a month, and takes anywhere from under five minutes to 45 minutes per report. It can't read images inside documents, so a scanned price list or a photographed quote won't count.

Who can use the Copilot Researcher agent?

The Copilot Researcher agent is available to business users with a Microsoft 365 Copilot licence and to Microsoft 365 Premium and Pro subscribers; free Copilot Chat users don't get it. Copilot Chat without a licence has limited agent access and no advanced agents.

PlanResearcherList price (checked 22 November 2026)
Copilot Chat on an eligible Microsoft 365 business planNoNo extra cost
Microsoft 365 Copilot Business (add-on)Yes, up to 25 queries a user a month$21 a user a month on a yearly subscription ($18 for the first year until 31 December 2026), or $25.20 a month on a monthly subscription
Microsoft 365 Premium (personal plan)Yes$199.99 a year or $19.99 a month at the regular price

For a business, the add-on is the normal route: the emails and files you want Researcher to compare sit on work accounts, and Premium works on a personal Microsoft account. Microsoft's pricing page also flags usage-based billing for some advanced AI features on top of the licence, and the business add-on comes with a one-month free trial that needs a credit card.

If you have a licence and Researcher isn't there, your admin may not have enabled it. Admins can block Researcher and Analyst in the Microsoft 365 admin centre, and individual users can't hide or switch Researcher off themselves. Turning Copilot agents back on when your organisation has disabled them walks an admin through the fix.

Where do you find Researcher in Microsoft Copilot?

Researcher sits under Agents in the Microsoft Copilot app: sign in at m365.cloud.microsoft, open Agents and select Researcher, or start from Home with the plus button, then Insights, then Research a topic.

  1. Sign in to the Microsoft Copilot app at m365.cloud.microsoft with your work account. Researcher is also in the Copilot mobile app on iOS and Android.
  2. Select Agents, then Researcher. If it isn't listed, try Agents, then More agents & plugins.
  3. Or, from Home, select the + in the chat box, then Insights, then Research a topic.
  4. Check the Web search toggle in Researcher's input box: on for a comparison that needs suppliers' public terms, off for one that should use only your own emails and files.
  5. Optionally, choose the model and mode with the model picker in the top right corner. Claude appears there only if your admin allows Anthropic models.

You can also type @Researcher inside an ordinary Copilot chat. In that case the model picker is disabled and the model is chosen automatically, which is fine for most supplier work.

How do you write a supplier comparison brief for Researcher?

A supplier comparison brief for Researcher names the suppliers, points to the emails and files holding their quotes, lists the criteria, and says which source wins when the web and a quote disagree. Specific briefs give Researcher less room to guess.

Microsoft's own advice for Researcher is to be specific, decide whether it should use workplace content, the web or both, and engage with its clarifying questions. For supplier work that translates into a brief with six parts. Here it is filled in for the worked example further down:

GOAL: Decide which motor-parts supplier we should use from January.
SUPPLIERS: Supplier A (current), Supplier B, Supplier C.

SOURCES:
- My emails: quotes from Supplier B (sent 6 October) and Supplier C
  (sent 9 October), and Supplier A's price increase notice (October).
- My files: "Top 20 parts 2026.xlsx" and Supplier A's last six invoices
  (PDFs in the Parts folder in OneDrive).
- The web: each supplier's published trade terms, delivery and returns pages.

COMPARE: price on our top 20 parts basket at our monthly volumes; delivery
runs per day, order cut-off times and delivery charges; credit terms;
returns and restocking fees; how warranty claims on parts are handled.

RULES: Quotes and emails from a supplier override its web pages. Cite every
figure. Mark each figure "from quote", "from invoice" or "from web".
Flag any web source older than 12 months. List anything you could not find.

OUTPUT: a comparison table, then monthly cost at our volumes
(80 orders a month, about 60 of them under $150, $400 a month of returns),
then open questions to ask each supplier before we switch.

Three lines in that brief do most of the work. "Quotes override web pages" stops a stale promotion page beating the price you were actually offered. "Mark each figure by source" turns the report into something you can check in minutes. And the volumes (80 orders, 60 under $150, $400 of returns) let Researcher turn per-order charges into monthly money, which is where supplier comparisons are usually won or lost.

Researcher may answer with questions before it starts. In this case it might ask whether price should outweigh delivery speed, and whether to apply Supplier A's announced increase to the comparison. Answer them in a line each ("price first, but a supplier with fewer than two daily runs is out"; "yes, use A's January prices"), or reply "Go ahead" if the brief already covers it. The answers steer the whole report, so they're worth the extra minute.

A garage compares three parts suppliers with Researcher

Take an illustrative four-bay car repair garage with five technicians, a service adviser and the owner. It spends about $6,800 a month on parts with its current motor factor, Supplier A, which has emailed notice of a 3% rise from January. Supplier B and Supplier C have both quoted by email on the garage's 20 most-used parts. The owner runs the brief above with Web search on, answers two clarifying questions, and gets the report back in about 20 minutes.

An illustrative extract of the report's comparison table, with the source labels the brief asked for:

CriterionSupplier A (current)Supplier BSupplier C
Top-20 basket, monthly$6,800 now, $7,004 from January (from email)$6,392, 6% below A today (from quote)$6,528, 4% below A today (from quote)
Delivery runs2 a day (from invoice notes)2 a day (from web)3 a day (from quote)
Delivery chargeNone (from invoices)$9.50 on orders under $150 (from web)None (from quote)
Credit terms30 days end of month (from invoice)30 days (from quote)30 days end of month (from quote)
ReturnsFree within 30 days (from web)Free within 14 days (from web)10% restocking fee after 7 days (from web)
Warranty claimsCredit on return of part (from web)Not foundNot found

The monthly cost section is where the report earns its keep. Supplier B looks cheapest on the basket, but its delivery charge applies to most of the garage's orders: 60 orders under $150 at $9.50 is $570 a month, taking B to $6,962. Supplier C has no delivery charge; even if every one of the $400 of monthly returns attracted its 10% restocking fee, that's $40, taking C to $6,568. Against Supplier A's January price of $7,004, B saves $42 a month and C saves $436 a month, or $5,232 a year.

Done by hand, the same comparison meant opening two quote emails, six invoices and three suppliers' websites, then building the sums in a spreadsheet: about two and a half hours for the owner, usually squeezed into an evening. With Researcher it was 10 minutes on the brief, 20 minutes of waiting while the bays kept working, and 30 minutes checking citations. The checking is not optional, as the next section shows, but the owner's time moved from collecting numbers to judging them.

If building the 20-part price basket in Excel is the slow part, try The Complete Excel Course with AI Integration on Coding Liquids.

How do you check Researcher's citations before deciding?

Check Researcher's citations by opening every source behind a figure you'll act on, confirming the number and its date there, and asking Researcher to list anything it couldn't source. Citations are a core part of Researcher's design, but a citation only proves where a figure came from, not that it's current.

  1. Open the money figures first. Basket prices, delivery charges, restocking fees and credit terms. Each should open to the quote email, the invoice or the supplier's own page.
  2. Read the date on every web source. A supplier's trade terms page from two years ago may describe a different deal.
  3. Look at what was searched. Microsoft shows the exact web search queries Copilot sent to Bing in the response's citations, so you can see whether it searched for the right supplier.
  4. Ask for the gaps. A follow-up such as "List every figure in your report that you could not source, and every criterion where a supplier's sources disagree" brings the weak spots to the top.
  5. Confirm with the supplier. Anything that decides the switch, such as a delivery charge or credit terms, goes into an email to that supplier before you sign.

The credit terms that came from an old promotion page

In the garage's first run, before the "quotes override web pages" rule was in the brief, the report gave Supplier C's credit terms as 60 days. The citation opened to a new-account promotion page on C's website, dated more than two years earlier. C's quote email, sent in October, said 30 days end of month. The owner noticed only because the cash-flow sum looked too good: an extra month's credit on about $6,500 of parts would have meant holding roughly $6,500 more cash in the business, and he'd started planning around it.

The fix was the rule in the brief plus the source labels, and a standing habit: any figure marked "from web" that affects money gets confirmed by email. Researcher had done what it was asked; the brief simply hadn't told it which source to trust. Microsoft's Researcher FAQ adds one more reason for care: with web search turned off, anything Researcher says about the outside world reflects the model's training data, which runs to 2024.

A follow-up prompt that finds the weak spots

Once the first report is in, one follow-up does more than any amount of rereading:

From your report, list every figure that decides which supplier is cheapest
over 12 months. For each, give the source, its date, and whether it came from
the supplier's own quote, an invoice or a web page. Then list the figures
you could not find at all.

An illustrative answer for the garage:

Deciding figures: Supplier B delivery charge $9.50 under $150 (web, trade terms page, updated this year); Supplier C basket $6,528 (quote, 9 October); Supplier A January basket $7,004 (email, price notice, October); Supplier C restocking fee 10% after 7 days (web, returns page, date not shown). Not found: warranty claim process for Suppliers B and C; whether Supplier B's delivery charge applies per order or per drop.

What to fix: the restocking fee has no date, so it goes on the list to confirm, and the "per order or per drop" question matters, because if B charges once per delivery run rather than per order, its monthly charge falls well below $570 and the comparison tightens. Microsoft's help doesn't say whether a follow-up counts against the 25 monthly queries, so budget as if it does and make each follow-up count.

Are confidential quotes safe when Researcher searches the web?

Confidential supplier quotes stay inside Microsoft 365 when Researcher searches the web: only a few-word query generated from your prompt goes to Bing, never whole emails or files, Microsoft says. The query carries no user name, domain or tenant ID.

Microsoft Learn's web search privacy page for Copilot applies the same rules to Researcher. The details that matter for supplier work:

  • The generated query is a few words informed by your prompt, not the prompt itself (unless the prompt is very short). It may be shaped by a document you explicitly reference, but the document isn't sent.
  • Your device's IP address is sent so Bing can localise results.
  • Prompts and responses stay within the Microsoft 365 service boundary under Microsoft's enterprise data protection. The generated search queries are handled under separate Product Terms commitments: not used to improve Bing, not used for advertising profiles, not used to train foundation models, and treated as customer confidential information.
  • Researcher follows the same data loss prevention (DLP) policies as Copilot, isn't trained on your organisation's data, and only reaches emails and files you already have permission to open.

Three switches control web search, from narrowest to widest:

ControlWhereEffect
Researcher's Web search toggleIn Researcher's input boxTurns web search off for your research
Web search switchMicrosoft Copilot app: Settings, Personalisation, AdvancedTurns web content off across your Copilot use; the choice persists across sessions and devices
Allow web search in Copilot policyAdmin, in the Cloud Policy service for Microsoft 365Turns web search on or off for users or groups; when off, users can't switch it back on

For a sensitive comparison, such as quotes given under a confidentiality agreement, split the work into two runs. Run one with Web search off and only your emails and files: the quotes, the invoices, the basket. Run two with Web search on and a prompt that names the suppliers and the public questions only (delivery terms, returns policy, warranty process), with no prices in it. It costs one extra query out of 25 and leaves nothing about your negotiation in the generated searches. How Microsoft 365 Copilot handles business data covers the wider privacy picture.

If you want help fitting Researcher into how your business buys and reviews suppliers, I offer one-to-one AI consultations for small businesses.

How long does a Researcher report take, and how many can you run?

A Researcher report takes under five minutes for a simple question and 10 to 45 minutes for a complex one, according to Microsoft's FAQ, and each user can run up to 25 queries a month.

Researcher spends its time retrieving and analysing deliberately, which is why Microsoft positions ordinary Copilot Chat for quick summaries and short replies and Researcher for multi-source questions that end in a report. Start a comparison, then go and do something else; on mobile you can check back from the Copilot app.

The 25-query ceiling shapes how a small firm should use it. A quick sum for the garage's owner, who also uses Researcher for other jobs:

Use in a typical monthQueries
Parts supplier review (first run plus one follow-up)2
Separate web-only pass for public terms1
Tyre supplier review, run every second month1
Equipment lease comparison (diagnostic machine)2
Customer or market questions the owner researches6
Total12 of 25

That leaves room, but only because the owner keeps quick questions ("what's B's phone number?") in ordinary Copilot Chat. Microsoft also says you can't yet control the length or detail of a Researcher report, and that PowerPoint and PDF output from Researcher is still to come, so plan to edit the report or use it to start a document rather than sending it as it is.

When is Researcher the wrong tool for a supplier comparison?

Researcher is the wrong tool when the comparison is really a spreadsheet job, when the quotes are images, or when a two-line question would do. Microsoft's own FAQ points spreadsheet-heavy work to the Analyst agent instead.

  • Price lists with hundreds of lines. Matching 400 part numbers across three price files is calculation, not research. Analyst handles Excel-style work better; using Copilot's Analyst agent on a sales spreadsheet shows how it works.
  • Scanned or photographed quotes. Researcher can't read images inside documents. Get text versions or type the figures into the brief.
  • Two quotes, one question. "Which of these two quotes is cheaper on these five items?" is a job for ordinary Copilot Chat or a side-by-side method such as comparing supplier quotes with AI, and it doesn't spend one of your 25 queries.
  • Background checks on a new supplier. Researcher can search the web, but checking who a supplier is before you sign is its own job, covered in checking a new supplier before you sign.
  • Custom rules. You can't change Researcher's behaviour or connect it to extra data through Copilot Studio; if your comparisons need a fixed scoring method every time, keep the method in your brief.

Which other supplier comparisons suit Researcher?

Supplier comparisons suit Researcher when the answer needs your own quotes and the suppliers' public terms together, and when the per-order details change the total. Two more illustrative businesses show the pattern.

A furniture maker choosing a hardwood timber merchant

A furniture maker buying about 40 cubic metres of oak a year has quotes from three timber merchants in email: $1,650, $1,580 and $1,720 a cubic metre, so $66,000, $63,200 and $68,800 a year on paper. The web side matters as much as the price: each merchant's published certification claims, kiln-drying standards, minimum orders and lead times. The brief tells Researcher to quote each merchant's own wording on certification and to flag any claim without a certificate number, because a sustainability line on a website is not a certificate. In the illustrative run, the cheapest merchant's certification claim had no number anywhere on its site and a 10-week lead time against 4 weeks for the others. At 40 cubic metres a year, the $2,800 saving against the first merchant had to be weighed against holding six more weeks of stock (about 4.6 cubic metres, or roughly $7,300 of timber sitting in the yard), and the owner asked for the certificate before going further.

An online clothing store comparing mailer-bag suppliers

An online clothing store sending 30,000 orders a year has three mailer-bag quotes: $0.19, $0.21 and $0.24 a bag, or $5,700, $6,300 and $7,200 a year. The brief asks Researcher to separate "the supplier says" from "an independent page says" in every row, because recycled-content percentages are exactly the claims that end up on the store's own product pages. The illustrative report showed the cheapest supplier needed a minimum order of 10,000 bags with a six-week lead time (about $1,900 paid up front, four months of stock), while the middle option allowed 2,500-bag orders in two weeks. For a small store watching cash, $600 a year more for the middle supplier bought back about $1,400 of cash tied up in bags, and the owner asked both for evidence behind their recycled-content figures before printing any claim.

From the Researcher report to a supplier decision

A report is not a decision. The garage owner turns it into a one-page scorecard, weighting what matters to a busy workshop, and scores each supplier from 1 to 5. The illustrative result:

Criterion (weight)Supplier ASupplier BSupplier C
Price after charges (40%)235
Delivery runs and cut-offs (25%)435
Credit terms (15%)434
Returns (10%)543
Warranty claims (10%)433 (to confirm)
Weighted score3.303.104.45

Supplier C wins clearly, but two cells were "not found" in the report, so the owner doesn't switch yet. He asks Copilot to draft an email to C confirming the four things that decide the switch: the basket prices held for 12 months, 30 days end of month credit, how a warranty claim on a failed part is credited, and whether the 10% restocking fee applies to parts ordered in error by C. Only when the reply confirms them in writing does the garage open the account, and Supplier A gets a polite email with the new prices it would need to match.

Keep the report, the scorecard and the supplier's written confirmation together in one folder. Microsoft's FAQ says Researcher sessions aren't directly available to admins or compliance tools by default, so if you need a record of why you switched, save it yourself; the report you edited and the email trail are that record.

Copilot Researcher for supplier comparisons: quick answers

Is the Researcher agent included with free Copilot Chat?

No. Researcher comes with a Microsoft 365 Copilot licence on a work account, and with Microsoft 365 Premium and Pro on a personal account. Copilot Chat without a licence has limited agent access and no advanced agents such as Researcher or Analyst. If you have a licence and still can't see it, your admin may have blocked it.

How many Researcher reports can I run each month?

Microsoft's Researcher FAQ gives a maximum of 25 queries per user per month for Researcher in Microsoft 365. A careful supplier comparison usually takes two or three runs, a first pass plus a follow-up or a separate web-only pass, so plan the month's research before spending queries on questions ordinary Copilot Chat could answer.

Can Researcher read a scanned price list or a photo of a quote?

No. Microsoft's FAQ says Researcher can't process images in input documents, so a scanned PDF, a photographed quote or a price list pasted as a picture won't be read. Ask the supplier for a text PDF or spreadsheet, or type the key prices into your brief, and say in the brief that those figures came from the supplier.

Can I stop Researcher from searching the web?

Yes. Researcher has a Web search toggle in its input box, and the Microsoft Copilot app has a Web search switch under Settings, Personalisation, Advanced that persists across sessions. Admins can also turn web search off for everyone with the Allow web search in Copilot policy, in which case users can't switch it back on.

Does Researcher send my supplier emails to Bing?

No. Microsoft says the query sent to Bing is a few words generated from your prompt, never your whole prompt (unless it is very short) and never whole emails, documents or uploaded files. It carries no user name, domain or tenant ID, though your device's IP address is sent so Bing can localise results.

Can I use Claude inside the Researcher agent?

Yes, if your admin allows Anthropic models in the Microsoft 365 admin centre. Inside Researcher, the model picker in the top right corner lets you choose the model and mode. If you call Researcher with @Researcher from an ordinary Copilot chat, the picker is disabled and the model is chosen automatically.

Further reads

Show 2 more related tutorials

Sources: Microsoft Support's Get started with Researcher; Microsoft Learn's What is Researcher Agent, the Researcher agent FAQ and Data, privacy and security for web search in Microsoft Copilot; Microsoft Support's How Copilot Chat works with and without a licence; Microsoft's Copilot business pricing page and Microsoft 365 Premium product page. All read 22 November 2026.

Ready to put AI on the jobs that eat your week?

No data team or big budget needed. In a 1:1 consultation I show small and medium-sized businesses and self-employed people which AI tools they can use, one workflow at a time.

Book a 1:1 call with me