AI Tools and Depreciation in 2026: What Indian Businesses Can Claim

Your AWS bill arrives on the 3rd of the month, your SaaS stack has quietly grown to eleven subscriptions, and your CA keeps asking whether the new AI licences are being capitalised or expensed. Most founders and SMB owners we work with have no clear answer, and the cost of guessing is real: either you under-claim deductions every year, or you over-claim and invite a notice. This post is about the second question, and it leans on the Income-tax Act, 2025, which replaces the 1961 Act with effect from 1 April 2026. Over the last two years we have seen Indian businesses roughly double their spend on AI-linked software, from LLM API credits and copilots to AI-enabled CRM add-ons and voicebot platforms. That spend sits in an awkward middle ground for tax purposes. Some of it is a clean operating expense, some of it behaves like a capital asset, and some of it is ambiguous enough that the classification depends on how the contract is written. We'll walk through how to tell which bucket each cost falls into, how to document it, and where the common errors happen. This is a practitioner's guide, not tax advice for your specific situation. Run the classification past your chartered accountant before you change how you book things.

Key Takeaways

• Pure subscription fees (SaaS, API usage billed monthly) are generally a revenue expense and deductible in the year incurred. They are not depreciable assets.
• Perpetual or long-term software licences you own and can use independently may qualify as an intangible asset eligible for depreciation, typically at 25% on a written-down value basis under the existing block framework.
• Custom AI models or software built in-house for your business should be separated into development cost (capitalise where criteria are met) and ongoing maintenance (expense).
• Always map each invoice to a contract clause. Tax classification follows the substance of the arrangement, not the vendor's marketing label.
• GST input tax credit on eligible AI subscriptions is separate from the income-tax deduction question, and the two should be tracked independently.
• Keep a register of AI tools with contract terms, renewal dates, and classification, so the books survive an audit without a scramble.

Why AI Tools Depreciation in India Income Tax Act 2025 Is Harder Than It Looks

The classic question for software was simple: is this a licence you bought outright, which can be depreciated as an intangible, or a service you rent, which is expensed? Cloud and AI tools have muddied that line. Vendors now offer annual prepaid plans, consumption-based pricing, enterprise agreements with minimum commitments, and bundled features that mix a licence with a service. Each structure can lead to a different treatment. The Income-tax Act, 2025 keeps the core principle that expenditure wholly and exclusively for business is deductible, and that intangible assets such as software acquired for business use can be depreciated. The mechanics of block-wise depreciation and the rates have carried over in substance from the earlier regime, so the practical question for your finance team is not "is there a new rule for AI," but "does this specific arrangement give us an asset or a service?"

The three questions that settle most cases

1. Do we own or control a usable asset after the contract ends? If you walk away with code, a model weight file, or a perpetual licence you can run on your own infrastructure, you are closer to an asset. If access stops when you stop paying, it is almost certainly a service.
2. Is the fee for access over time or for a discrete deliverable? Monthly seats and API usage are time-based. A one-time build of a model pipeline delivered to you is closer to a capital cost.
3. Would the cost recur at the same level to keep the business running? Recurring costs to maintain operations lean towards expense. Large one-time outlays that create a lasting capability lean towards capitalisation.

Answer these three questions for every AI-related invoice and you will classify about 90 percent of your spend correctly. The remaining 10 percent is where a CA's judgement matters.

How Do You Treat AI Software Subscriptions Under the New Act?

Most SMB AI spend is subscription-based, and for these the answer is straightforward. A ChatGPT Team or Enterprise seat, a Claude or Gemini API bill, a Jasper or Copy.ai subscription, an AI feature bundled into a Zoho or HubSpot plan: each of these is a periodic payment for access. Treat it as a revenue expense in the period it relates to.

The subtlety is timing. If you prepay an annual plan in March for a full year, the expense is spread over the period of benefit. For a company that closes its books on 31 March, that means you can deduct the portion relating to the current year and carry the rest as a prepayment. Many small firms simply expense the whole annual invoice in the year paid, which is acceptable for small amounts but becomes a finding when the sums are large.

A worked example: a 15-person logistics company in Gurgaon

Consider a logistics firm in Gurgaon with fifteen staff, running dispatch operations and customer support. Over FY 2025-26 they adopted an AI routing assistant at ₹38,000 per month, an LLM API for customer chat at roughly ₹22,000 per month on average usage, and a seat-based AI meeting-notes tool for six managers at ₹1,800 per seat per month, or ₹10,800 monthly.

Their annual AI subscription spend came to about ₹9.6 lakh (₹38,000 + ₹22,000 + ₹10,800 = ₹70,800 per month, times 12 = ₹8,49,600, plus a ₹1.1 lakh annual true-up on API overage, rounding to roughly ₹9.6 lakh). All of this is recurring access, so it is an operating expense and fully deductible in the year incurred, subject to the wholly-and-exclusively-for-business test. The firm's error was not the classification but the documentation. The meeting-notes seats were partly used by one manager for personal calendar management. Splitting that usage and disallowing the personal share, supported by a usage-based allocation, is the kind of detail that protects the rest of the claim.

When Does an AI Licence Become a Depreciable Asset?

This is where most confusion sits. A perpetual licence to an AI-powered analytics package, installed on your own servers, with no recurring fee beyond maintenance, is software you own. Its cost is a capital expenditure. Under the depreciation framework that carries over into the new Act, computer software is typically placed in a block with a rate of 25 percent on a written-down value basis, which is the same treatment historically applied to software under the earlier schedules.

Be careful about what you are actually buying. A fine-tuned model that your vendor hosts and charges you to query is still a service. A model weight file you download and run on your own GPU instance, under a licence that survives termination, begins to look like an asset. The distinction is contractual and operational, and you should read the licence before you read the marketing deck.

Custom AI development: separate the build from the run

Many Indian SMBs commission custom software or AI integrations from agencies. A recommendation engine built for an e-commerce client in Pune, or a document-extraction pipeline built for a Chennai logistics firm, typically has three cost types. The design and development phase, where a working system is created and delivered to you, is generally capitalisable as an intangible asset once it is ready for use. Hosting and cloud compute during operation is an expense. Ongoing support, bug fixes and minor changes are an expense. Training and change-management costs for staff are an expense. The mistake we see most often is capitalising everything in a single invoice because it is labelled "AI implementation." Ask your vendor for a breakdown by phase before you book it. A vendor who cannot provide this is making your audit harder, whatever their intentions.

Pro Tip: Ask vendors for milestone-based invoicing with separate line items for licence fees, development, hosting, and support. Vendors who bill a lump sum for an "AI platform" give you no basis to split the cost, and a tax officer will fill that gap with their own assumptions, usually unfavourable ones.

Cloud, GPU Compute and AI Infrastructure: Which Bucket?

Cloud infrastructure costs for running AI workloads are almost always operating expenses. Compute hours on AWS, GCP or Azure, storage, data transfer, managed AI service calls and GPU instances rented by the hour are all services consumed over time. The fact that they support an AI product does not make them assets of your business.

The capitalisation question arises when you buy hardware. An on-premise GPU server bought outright is tangible plant and machinery, and it depreciates under the appropriate block for computers, which has historically been 40 percent on a written-down value basis. That rate is substantially higher than the software rate, which is why the hardware-versus-cloud decision is partly a tax-planning question as well as an operational one.

Here is a comparison of common approaches for an SMB deciding how to host AI workloads.

Option	Typical cost pattern	Tax treatment	Best suited for
Public cloud (AWS, GCP, Azure) on-demand	Pay per hour or per call, variable monthly bill	Revenue expense, deductible when incurred	Startups and workloads with uneven demand
Cloud reserved or committed-use plans	Discounted rate for 1 or 3 year commitment, prepaid or monthly	Expense over period of benefit; prepaid portion carried forward	Steady production workloads with predictable usage
Managed AI APIs (OpenAI, Anthropic, Google Gemini API)	Usage-based billing, often in USD with GST considerations	Revenue expense; track forex impact separately	Feature-led products and experimentation
On-premise GPU server purchased outright	Large upfront capex plus power, cooling, maintenance	Tangible asset; depreciation on plant and machinery block	Regulated data, consistent heavy workloads, mature businesses
Colocation with owned hardware	Capex for hardware, monthly rack and bandwidth fees	Hardware capitalised; colocation fees expensed	Firms that want owned hardware without running a data room

In practice, the financially sensible choice for most SMBs is public cloud or managed APIs until workloads are predictable enough to justify committed plans. If you are weighing a move from on-premise to cloud, our cloud migration and managed services team at eDarpan's cloud services practice regularly models both the operational and tax implications before any hardware is retired.

A Step-by-Step Walkthrough: Building an AI Tools Register

The single most useful thing an SMB can do before its next audit is to maintain a register of AI tools. Here is a process you can run yourself or brief to your finance team.

1. List every tool in use, including free tiers and personal accounts used for work. Ask department heads, not just IT. Marketing and customer support often adopt tools without informing finance.
2. Capture the contract terms for each. Record vendor, plan, billing currency, billing frequency, renewal date, whether the licence is perpetual or subscription, and whether there is a minimum commitment.
3. Assign a classification. Use the three questions above. Mark each tool as expense, capitalise, or split, with a one-line reason.
4. Attach the evidence. Link each tool to its invoice, contract, and a short business-purpose note. For shared seats, record the allocation basis.
5. Track GST separately. Note whether the vendor is registered under GST in India and charges IGST or CGST and SGST, or whether it is a foreign supplier subject to reverse charge. Input tax credit eligibility depends on this, and it is independent of the income-tax treatment.
6. Flag currency exposure. If you pay a US-dollar invoice, record the exchange rate used for booking. Rupee movements change the cost in your books without any change in what you bought.
7. Review quarterly. Remove cancelled tools, add new ones, and check that renewal dates are visible to the person who approves spend.

A spreadsheet is enough for most companies under fifty people. The value is not the tool but the habit: when the statutory auditor asks about a ₹4 lakh invoice, you open one row and answer.

What About GST on AI Tools?

GST and income-tax treatment are two separate questions, and conflating them causes errors. A foreign AI vendor that does not have an Indian presence may require the recipient to pay GST under the reverse charge mechanism on imported services, subject to the rules in force. Indian-registered vendors charge GST that you can typically claim as input tax credit if the expense is used for business and the vendor has filed correctly.

The practical risk is that the credit is lost because the vendor did not file its GSTR-1 or GSTR-3B on time, which is a recurring problem in the SaaS sector. Check your GSTR-2B regularly, and if a vendor's invoice is missing, raise it before the credit lapses. For businesses that also sell online and are thinking about how their registration footprint should grow, our note on multi-state GST registration for e-commerce sellers covers when expansion makes sense.

Common Mistakes Indian Finance Teams Make With AI Spend

Beyond the classification questions above, we see a few errors repeatedly.

Booking pilots as capital projects

A three-month AI pilot that never reaches production is usually an expense. The temptation is to capitalise the whole project in the hope it will eventually be used. Unless the output is an asset you will use in the business, the cost should be expensed. A pilot that is abandoned leaves an asset on the books for a system that does not exist, which is awkward to explain.

Ignoring personal use

Founders often run AI subscriptions on company accounts for personal productivity. The deduction is only for business use, and a clear allocation is needed. A simple rule, such as 80 percent business based on a usage log, is more defensible than an unsupported 100 percent.

Forgetting annual true-ups

Usage-based AI billing can produce a large invoice at year-end when overage charges land. Accruing an estimate each month avoids a distorted profit picture and a surprise deduction timing issue.

Relying on vendor invoices that do not state the plan

An invoice reading "AI services ₹2,40,000" gives you nothing to classify. Insist on invoices that name the product, the plan, the billing period and, where relevant, the licence type.

How Should an SMB Plan AI Adoption Costs in Its Budget?

For finance-minded decision-makers, the goal is to forecast AI spend the way you forecast rent or payroll, with a clear view of what is recurring and what is one-time. We recommend splitting the budget into three lines: committed recurring subscriptions, variable usage (API and cloud consumption), and project capex for anything you intend to own. Review the mix each quarter. A useful planning heuristic is to set a ceiling on variable usage and a trigger for review. When API spend crosses a set threshold, someone must confirm the use case still justifies it. This keeps experimentation from quietly becoming a fixed cost nobody approved. For businesses that are still choosing their workplace footprint alongside these decisions, the trade-offs in office leasing versus a virtual office are worth weighing in the same budget cycle.

Real Costs for Common Indian AI and Messaging Stacks

To make the planning concrete, here is how a few popular categories typically behave in rupee terms. Figures vary by plan, usage and vendor pricing changes, so treat these as planning ranges rather than quotes.

• Seat-based AI assistants for a team of ten commonly run from roughly ₹15,000 to ₹40,000 per month depending on tier and whether enterprise controls are required.
• LLM API usage for a customer-support chatbot handling a few thousand conversations a month often lands between ₹10,000 and ₹60,000, with wide variance driven by context length and model choice.
• AI voicebot platforms are usually priced per minute of call handled, and the costs scale quickly with volume. Our team's AI voicebot service is built around modelling this cost against expected call volume before a deployment starts.
• WhatsApp-based AI workflows depend on conversation-based pricing from Meta and the BSP layer on top. If you are integrating these, see our WhatsApp Business API practice for how we structure the commercial model.

Implementation Checklist for the Next 30 Days

If you are reading this with an audit or budget review approaching, here is a practical sequence.

1. Pull the last twelve months of payments to AI vendors from your bank and card statements, and match each to an invoice.
2. Build the register described above and classify each line.
3. Send the list of capitalisation candidates to your CA with the contracts attached, and ask for a written view.
4. Ask every vendor in use for a GST-compliant invoice naming the plan and billing period.
5. Set a quarterly review date and an owner for the register.
6. Decide on the hosting question for the next financial year: cloud on demand, committed cloud, or owned hardware, using the table above as a starting point.

Frequently Asked Questions

Can I claim depreciation on ChatGPT or Claude subscriptions in India?

Monthly or annual subscriptions to these services are generally treated as operating expenses, not depreciable assets, because you do not own a usable asset when the subscription ends. The full amount relating to the period can be deducted as a business expense, provided it is used wholly and exclusively for business. Prepaid annual plans should be spread over the period they cover.

Is AI software bought outright eligible for depreciation under the Income-tax Act, 2025?

Yes, if you acquire a licence you own and can use independently of the vendor, such as a perpetual licence to software installed on your systems. Such software is typically treated as an intangible asset and depreciated, with the rate historically applied to computer software being 25 percent on a written-down value basis. Confirm the current rate and block treatment with your chartered accountant, since rules can be refined through rules and circulars.

How do I classify a custom AI model built by an agency?

Split the invoice into development, hosting, and support. The development cost that produces a working asset delivered to you is generally capitalisable once it is ready for use. Hosting and support are operating expenses. Ask the agency for a phase-wise breakdown before you book anything, and keep the statement of work on file.

Is GPU cloud compute a capital expense or an operating expense?

Renting GPU compute on AWS, GCP, Azure or a similar provider is an operating expense, because you consume a service over time. Buying a GPU server outright is a capital expense for tangible plant and machinery, which depreciates separately. The choice between these is both a cost question and a tax-timing question.

Does GST input credit on AI tools depend on how I depreciate them?

No. GST input tax credit eligibility depends on whether the supply is used in the course of business and whether the vendor has correctly declared and paid the tax, not on whether the item is depreciated or expensed. Track GST credit through GSTR-2B separately from your income-tax classification.

What records should I keep for AI tool expenses in case of a tax notice?

Keep the vendor invoice showing the product and plan, the contract or terms, proof of payment, a business-purpose note, the allocation basis for any shared or part-personal use, and your classification memo. A tool register that links these together is the fastest way to respond to a query.

Should a small business hire help to set this up?

For most SMBs, a one-time review with a CA plus a simple register is enough. If you are also modernising your infrastructure, the classification question becomes tied to architecture decisions, and it helps to have both perspectives in one conversation. eDarpan's IT consulting team works alongside finance advisers on exactly this kind of decision, and you can reach us through the contact page to discuss your setup.

Closing Thoughts

AI tools depreciation India Income Tax Act 2025 is not a question with one universal answer, but it is a question with a repeatable method. Separate access from ownership, separate build from run, track what you buy in a register, and let the contract, not the marketing, decide the classification. Businesses that do this consistently spend less time arguing with auditors and more time deciding which AI investments actually earn their place.

If you are planning AI adoption for the coming financial year and want help modelling the cost and the tax treatment together, eDarpan can help you map the options against your budget. Learn more about who we are and how we work, or explore our full range of technology services. We are glad to walk through your numbers before you commit.

Image credit: Out of business by kevin dooley via flickr (BY 2.0), sourced through Openverse.