Automated Managed Service Provider: Hardware Refresh Plans From Asset Lists

Export each client's assets with device type, model, purchase date, warranty end and operating system, add vendor end-of-support dates, then ask AI to sort every device into replace now, replace next year or later. You check the groups against lifecycle pages and ticket history, and AI builds the budget table from your distributor's prices.

The grouping is the easy part. The value, and the risk, sit in the inputs: an asset list with blank warranty dates produces a confident wrong plan, and a model asked for end-of-support dates from memory will supply some that are wrong. Keep the dates in a lookup you build from vendor pages, give the model your bucket rules in writing, and recompute every budget line it produces.

Follow me on Instagram@sagnikteaches

What a refresh plan has to answer for the client

A business owner reading a refresh proposal has three questions. What is at risk now? What will it cost over the next twelve months? What can wait? Everything in the plan should answer one of those. Your own technical grouping (age bands, model families, firmware) is a means to that, not the output.

Connect on LinkedInSagnik Bhattacharya

It helps to name the reasons a device lands in each bucket, because the reasons are what you defend in the review meeting:

Subscribe on YouTube@sagnikteaches

• Replace now: the vendor no longer issues security updates for the operating system or firmware, the device is out of warranty and failing, or it cannot run something the business needs.
• Replace next year: support ends within the next twelve to eighteen months, warranty ends within twelve months, or the device is older than your standard cycle (many shops use four years for laptops, five to six for servers and network gear; use your own).
• Later: inside support, inside warranty or comfortably within your cycle, with no ticket pattern suggesting trouble.

Write those rules down once, in your own words, before you involve AI. The model applies rules; it should not invent them.

Step one: export the asset list with the fields that decide a refresh

Pull one CSV per client from your RMM, PSA or documentation tool. The columns that matter are listed below. If your tool cannot export one, add it by hand for the few devices in question, not for the whole estate.

Leave out names, email addresses, IP addresses and serial numbers before any of it goes into an AI tool. The asset ID is enough to map the answer back. If you need serials to look up warranties, do that step in the manufacturer's own tool, not in the chatbot.

Step two: clean the list before AI sees it

Most asset lists are untidy in predictable ways. Spend fifteen minutes on a pass like this, which you can partly hand to AI:

1. Standardise model names ("Latitude 5420", "Dell Lat 5420" and "5420 laptop" become one).
2. Remove retired devices still sitting in the export.
3. Flag duplicates, such as one laptop appearing under two asset IDs after a rebuild.
4. Fill or flag every blank warranty and purchase date.
5. Check that the user group is present for every device.

A prompt that works for step one:

Below is a CSV of device models from one client's asset list.
Group rows that are the same model written differently and give me a two-column table:
original text, standard model name. Do not guess. If a row is ambiguous, put "check" in the second column.
Do not add or remove rows.

Illustrative sample output: of fifty model strings, forty-three map cleanly; seven are marked "check" because they say only "HP laptop" or "old Dell". Those seven are the first thing to visit or ask about. The model's honesty about ambiguity is what you are paying for; if it fills them in confidently, your prompt needs tightening or the model needs swapping.

Step three: take lifecycle dates from the vendor, not from memory

Lifecycle dates are the facts the whole plan hangs on, and the place where confident AI answers go most wrong. Build a small lookup table yourself from each vendor's lifecycle page and give it to the model as the only source. Three dates worth knowing as examples, all from Microsoft's own pages:

• Windows 10 support ended on 14 October 2025. Organisations can buy Extended Security Updates through volume licensing; Microsoft's programme page lists $61 per device for year one, with the price doubling each consecutive year, for a maximum of three years. Years are cumulative, so a device enrolled in year two pays for year one too, and only Windows 10 version 22H2 is eligible. It is a paid bridge, not a plan.
• Windows Server 2016 reaches the end of extended support in January 2027.
• Windows Server 2019 reaches the end of extended support in January 2029.

Re-read those pages before you use the figures in a client document; I fetched them in November 2026 and vendors revise pages. For hardware, the same approach applies: use the manufacturer's end-of-life or end-of-support notices for firewalls, switches and access points, and note the URL and the day you checked in your lookup sheet.

Structure the lookup like this, so the model has nothing to invent:

Anything not in the table must come back as "not in lookup", never as a date. This single rule prevents the worst error in the whole workflow.

Step four: the sorting prompt and what comes back

Now give the model your rules, the lookup table and the cleaned list, and ask for a classification with the reason for each row.

Rules (mine, apply exactly):
- REPLACE NOW: OS or firmware past end of support in the lookup; or out of warranty AND 6+ tickets in 12 months.
- REPLACE NEXT YEAR: support ends within 18 months of 1 November 2026; or warranty ends within 12 months; or age over 4 years (laptops/desktops), 6 years (servers, network).
- LATER: everything else.
Use only the lookup table for support dates. If a date is missing, write "not in lookup".
For each device return: asset ID, bucket, one-line reason citing the rule, and any data gap.
Return a table. Do not suggest products or prices.

Illustrative output for a twenty-two-user client with thirty-eight devices: eleven laptops still on Windows 10 land in REPLACE NOW because support has ended; one server on Windows Server 2016 lands in REPLACE NEXT YEAR because extended support ends in January 2027; two switches and the firewall land in LATER; three devices come back "not in lookup" because the model column was blank.

What you fix: check that the "next year" cut-off was applied to the right date, since models sometimes miscount months; check that a device with unknown warranty did not fall into LATER by default; and look at any device the PSA shows as a finance or director machine, which you may want to move up regardless of age.

Then look hard at the server. By the letter of the rules it is next-year work, and the model applied them correctly. But extended support for Windows Server 2016 ends in January 2027, about ten weeks after the plan date, and a server replacement rarely fits in ten weeks once you count the quote, the order, the build, the data move and a weekend cut-over. The rules, not the model, were wrong. Add a line such as "support ends within six months of the plan date: REPLACE NOW" and rerun. That one change moves the file server up a bucket, and it is exactly the kind of judgement the model will not make unless you write it down.

Step five: test the buckets against tickets and client plans

Lifecycle and age are blunt. Two more inputs change the answer.

Ticket history. A three-year-old laptop with nine tickets in a year is a worse bet than a five-year-old with none. Group your PSA tickets by asset and add the count as a column. The tutorial on how small IT support businesses use AI to resolve tickets faster shows how to keep ticket data tidy enough for this.

Client plans. Ask the client contact two questions: are you hiring or letting people go in the next twelve months, and are you moving premises or to cloud services? A planned office move turns a "later" network refresh into a "do it with the move" job. Put the answers into the prompt as notes and rerun only the affected rows.

A realistic mistake and how it showed up

Say a client's asset export lists a pile of laptops with purchase dates from the day the RMM agent was installed, not the day the laptops were bought. Everything looks two years old, so the plan puts nearly all of it in LATER. The error surfaces when the client asks why a three-year-old-looking laptop still has a Windows 7 sticker. The fix is the cleaning pass: compare install date against manufacture date from the warranty lookup, and use the earlier. Add "date source" as a column so the plan says which kind of date it used.

Step six: build the budget table with your prices, not the model's

Ask AI to do the arithmetic and the formatting, and give it the prices. A short price table from your distributor, with the date you took it, is the input.

Price list (my distributor, taken 3 Nov 2026, ex-tax, illustrative):
Standard laptop, 16 GB: 780
Desktop, small form factor: 640
Server, entry rack: 3,900
24-port managed switch: 420
Firewall appliance with 3-year licence: 1,150
Build a budget table from the classification below.
Columns: bucket, item, quantity, unit price, line total. Add subtotals per bucket and a grand total.
Add a 10% contingency line and an 8% deployment-hours line. Show your sums.

Illustrative result: eleven laptops at 780 is 8,580; the server at 3,900 sits in next year; two switches at 420 is 840; contingency and deployment lines are calculated on the bucket totals. Replace-now total 8,580 plus 10% contingency and 8% deployment gives 10,124.40; next-year total 3,900 plus the same loadings gives 4,602; later is 840 plus loadings, 991.20 if the client chooses to bring it forward. Once the six-month rule moves the server, replace-now becomes 8,580 plus 3,900, or 12,480 before loadings and 14,726.40 after them, and the next-year bucket shrinks by the same amount. All figures are invented for the example and use no real client.

Always recompute two lines by hand. If the model's multiplication or percentage is wrong, the table is unusable and you should say so in the prompt. Spreadsheet formulas are safer for the final version: ask the model to produce the layout, then put your own formulas in the cells.

A sorted extract: what the finished table looks like

Here is an illustrative extract from the twenty-two-user client, after the model has classified the rows and you have corrected it. Asset IDs are internal, and no real names or serials appear.

Notice the last two rows. The firewall is placed in LATER but carries a data gap, because nobody has checked the manufacturer's end-of-life notice yet; until someone does, the plan should say "provisional". The unidentified laptop stays out of the budget until it is looked at, and the summary tells the owner it exists. A plan that shows its gaps earns more trust than one that looks finished.

Servers and network gear need a different pass from laptops

Laptops can be rolled out in batches and failures are annoying but contained. Servers, firewalls and core switches fail all at once, so the rules should differ.

• Servers: weigh support dates, but also the recovery plan. Ask when the last restore test happened. A server with a tested restore and two more years of support can wait; one with no restore test cannot, whatever its age.
• Firewalls: subscription licences (for threat updates and similar) often end on a date separate from the hardware warranty. Record both dates in separate columns, because the licence date is usually the one that bites.
• Switches and access points: these tend to be replaced as a group during a network project. Sort them individually, then ask AI to propose a single "do together" line in the budget with the reasoning, and check that the grouping makes sense on site.
• Printers and peripherals: leave them out of the model run unless the client asks, and give a flat allowance instead.

Say a nursery has one small server holding children's records, backed up nightly to a cloud service. The restore has never been tested. Even with three years of support left, that is a replace-or-prove decision this quarter: either run a test restore and record the result, or move it up in the plan. That reasoning goes in the notes column; the model only follows it if you tell it.

The approval message to the client

Once the summary is agreed, the client needs a clear ask. AI can draft it from the two decisions, and you send it from your own account.

Draft a short email to [first name] confirming two decisions from the refresh plan:
1) approve replacing 11 laptops and the file server this quarter, estimated total 14,726.40 including contingency and deployment;
2) choose between a paid one-year extension for 3 remaining older machines or replacing them in the spring.
Say prices were taken on 3 November and may change by the order date.
Keep it under 120 words and offer a call.

Illustrative output is a tidy four-sentence email. What you fix: it will often write "we recommend" in places where you have not given a recommendation, and may describe prices as "fixed". Delete both. A client who reads "fixed" will hold you to it.

Three scenarios to test the plan with

Before you send anything, run these quick checks against the finished plan.

1. The cash-flow view. Spread the replace-now total across the months you will actually deploy. Many small clients prefer three instalments to one invoice.
2. The paid bridge. For devices that cannot run the current operating system, price the vendor's paid extended-update programme for one year against replacing them. Where the per-device bridge cost is a large fraction of the new device, replace; where only a handful of old machines remain and a swap is due anyway, a bridge may be sensible. This is a business decision for the client, with your recommendation.
3. The failure test. Ask what happens if the oldest server dies tomorrow. If the answer is two days' downtime and no tested restore, that device's priority rises whatever the date says.

This is the kind of reasoning that belongs in the quarterly meeting, and the tutorial on writing quarterly business reviews with AI shows where to put the refresh plan in the pack.

Presenting it: one page the owner reads in two minutes

The plan the client sees is not the spreadsheet. It is a one-page summary built from it. AI is good at this last step, given the table and instructions to keep to what is in the table.

Write a one-page summary for the owner of a 22-person business.
Use only the figures in the table below. Three headings: What needs replacing now and why;
What to budget for next year; What can wait. Plain language, no jargon,
explain "end of support" in one clause. End with two decisions I need from them.
Do not add any figure that is not in the table.

Edit the result: remove any claim about risk that you cannot back with a lifecycle page, and make sure the two decisions are real choices (approve the replace-now spend; choose between paid bridge and replacement for the remaining Windows 10 machines).

Where small MSPs trip up on this workflow

Treating the PSA as the truth

The asset list is only as good as the last person who updated it. Walk a sample of ten devices on site or via remote tools each quarter and compare against the list.

Sending the client your raw list

Clients do not want an export. They want a decision and a budget. The raw list is your evidence, shared on request.

Different examples for different clients

A specialist food importer with one warehouse server and a few handheld scanners has different priorities from a childcare nursery with a handful of tablets and strict data duties, and a kitchen fitter's laptops spend their lives in vans. A craft brewery supplier may care most about the label printers and the till. Put the role of each device in the list so the same prompt gives sensible answers across very different clients.

Letting AI choose products

Do not ask it to recommend models. Your distributor's current lines and your standard builds decide that, and a model's suggestions will often be discontinued or mismatched with your management tools.

A repeatable quarterly rhythm

1. Export fresh asset data per client (ten minutes each).
2. Run the cleaning prompt and resolve "check" rows.
3. Update the lifecycle lookup from vendor pages and note the date.
4. Run the sorting prompt; review exceptions.
5. Rebuild the budget table with current prices.
6. Generate the summary, edit, and attach it to the quarterly review.

If this is the first workflow you are standardising, five AI workflows for an MSP and their set-up time shows what to automate next. With the prompts saved, a twenty-device client takes about forty minutes end to end, the first time perhaps twice that. Those are my estimates for planning, not measured figures; time your first three clients. For choosing where the asset data should live, the comparison of ConnectWise AI agents and HaloPSA AI is a good next read, and the wider AI guide maps related tutorials.

Questions MSPs ask about AI-built refresh plans

Can I paste a client's asset list into a chatbot?

Strip anything that identifies people or networks first: usernames, serial numbers if not needed, IP addresses and hostnames that reveal structure. Use a business plan with terms you have read, and tell the client in your agreement that AI tools help prepare reports. Keep the full export inside your own systems.

How accurate are AI-estimated replacement prices?

Treat them as placeholders. A model has no live price list, so give it your distributor's current line prices as a table and ask it only to multiply. Anything it fills in from memory should be marked for checking before the budget reaches the client.

Which lifecycle dates should I trust?

Only the vendor's own lifecycle page, read in the month you write the plan. Record the date, the page and the day you checked it in the sheet. Warranty dates come from the manufacturer's lookup using the serial number, not from the purchase date.

How often should a refresh plan be redone?

Review it every quarter alongside the business review, and rebuild it fully once a year. A new lifecycle announcement, a failed device cluster or a client acquisition are also reasons to rerun it sooner.

Further reads

• Best AI Help Desk Tools for Small MSPs (2026) — Help desk tooling that feeds your asset records.
• AI Receptionist for an MSP: Cost and After-Hours Ticket Logging — Clean ticket intake gives you better failure evidence.
• Is AI Ticket Automation Worth It for a Small IT Support Shop? — Decide whether more automation pays for a small shop.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — A readiness check before adding more AI tasks.
• AI IT Documentation for MSPs: Client Runbooks From Screen Recordings — Client runbooks that record what each device does.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Learn lifecycle pages for Windows Server 2016 and Windows Server 2019, and the Microsoft Learn page on the Windows 10 Extended Security Updates programme, all read 4 November 2026; manufacturer warranty lookup tools.