Data annotation and AI training companies in India that label images, transcribe audio, rate model outputs or write training data for foreign AI firms are exporting services, and a business doing this files a monthly EDF with its AD bank. The industry's distinctive problem is not the form itself but usage-based billing that keeps changing after the invoice: tasks get rejected, re-done or credited, and the EDF must stay consistent with what the client finally pays. Handle that with clean invoicing rules and the EDF becomes a ten-minute monthly job.
Key takeaways
- One EDF per month covering every export invoice, due within 30 days from the end of the invoice month.
- Mode of delivery is Internet. Mode of realisation is usually Others.
- Usage-based invoices should be raised on accepted volume where possible, so fewer credit notes follow.
- Quality rejections after the EDF is filed are handled with credit notes and a written note to your bank, not by editing the EDF.
- SAC and purpose codes depend on how the work is framed (IT-enabled data services versus research support). Confirm with your CA and bank.
The work, and why it is a service export
Annotation shops do a wide range of work: bounding boxes and segmentation for computer vision, transcription and audio tagging, text classification, ranking and rating model responses, red-teaming prompts, writing domain-specific question-and-answer pairs, and evaluation of model outputs by subject experts. The output is delivered online, usually directly into the client's labelling platform.
When the client is abroad and pays in foreign currency, this is an export of services. If you operate as a company, LLP, firm or proprietorship, assume the EDF applies. Individual annotators working directly for foreign platforms are a separate question; RBI said on 7 October 2026 that individuals providing services abroad are not required to report, with FAQs to follow.
Billing models and their EDF footprint
| Billing model | How the invoice is built | EDF watch-out |
|---|---|---|
| Per task or per label | Accepted units x unit rate | Raise on accepted units to avoid later credit notes |
| Per annotator hour | Hours from the client's platform or your timesheets | Agree whose hours count before invoicing |
| Monthly managed workforce | Fixed fee for a team of N annotators | Straightforward; one row a month |
| Pilot or proof of concept | Fixed fee for a test batch | Small invoice; fits the ₹10 lakh declaration route |
| Expert evaluation (domain specialists) | Per hour or per evaluated item, higher rate | Describe accurately; may suit a different SAC |
The cleanest discipline is to invoice only after the client's quality review is complete for the period. If your contract makes you invoice on submitted volume and the client then rejects 8%, you are setting yourself up for a credit note every month.
Worked example: a month with a quality credit
Tessellate Labelworks Pvt Ltd, Coimbatore, works for three AI companies. Its March 2027 invoices:
| Sr. | Recipient | Country | Invoice | Currency | Amount | Description |
|---|---|---|---|---|---|---|
| 1 | Parallax Vision Inc | USA | TLW/27/M03-1 | USD | 18,420 | Image segmentation, 92,100 accepted masks |
| 2 | Corvus AI Ltd | UK | TLW/27/M03-2 | GBP | 9,600 | Managed rating team, 12 evaluators, Mar 2027 |
| 3 | Nordlys Language AS | Norway | TLW/27/M03-3 | USD | 2,750 | Pilot: Hindi-English audio transcription, 120 hours |
On 18 April, after the March EDF has gone to the bank, Parallax's audit rejects 3,000 masks from invoice 1 (USD 600). Tessellate issues credit note TLW/CN/27/004 for USD 600.
What it does:
- Leaves the March EDF as filed. It was correct when submitted.
- Writes to its AD bank, attaching the credit note, explaining that USD 17,820 rather than USD 18,420 will be realised against invoice 1.
- Asks the bank how it wants the reduction reflected in EDPMS.
- Records the change in its EDF register.
If the credit note had been issued before the EDF was filed, Tessellate could simply have shown NRV of USD 17,820 in that row, with the credit note referenced in remarks. Our guide on credit notes, discounts and short payments covers both timings, and the credit note generator produces the document.
The March 2027 EDF is due by 30 April 2027.
Codes: frame the work honestly
Annotation sits awkwardly between categories. Some firms treat it as an IT-enabled data service, others as research support for the client's machine learning work, and expert evaluation can look like consulting.
| If your work is mainly... | SAC direction (confirm with CA) | Purpose code direction (confirm with bank) |
|---|---|---|
| Data processing and labelling inside client platforms | IT services group, such as 998319 other IT services | Software or IT-related codes such as P0807 may be discussed |
| Supporting the client's R&D programme | Research services codes | P1008 research and development services |
| Domain expert review and advisory | Depends on the domain | Depends on the service |
Do not pick a code because it seems to give a better tax outcome. Pick the one your contract and deliverables support, and keep it consistent across invoice, FIRA and EDF. The SAC code finder is a sensible starting point.
Realisation and payment practicalities
AI clients usually pay by wire or through payment platforms. Whatever the route, the money should land in your AD bank account with a FIRA or bank advice you can tie to an invoice. Realisation is due within 9 months from each invoice date.
Watch these two issues:
- Netting: a client that deducts credits from the next payment instead of paying invoices in full creates matching headaches. Ask for payments referenced invoice by invoice.
- Pilots: small pilot invoices are easy to forget once the main contract starts. Use the periodic ₹10 lakh realisation declaration to close them in bulk; our realisation declaration tool builds it.
GST and LUT
Annotation and AI training work for foreign clients is generally an export of services when the five conditions in Section 2(6) of the IGST Act are met, so it is zero-rated. Export under LUT and print "Supply meant for export under LUT without payment of IGST" on each invoice. The EDF is a FEMA declaration and does not change GST or income tax. This is general information, not legal or tax advice.
Frequently asked questions
Do data annotation companies need to file the EDF?
Yes, if they are businesses exporting services. A proprietorship, firm, LLP or company that invoices foreign clients for annotation or AI training work should file one EDF a month covering that month's export invoices.
How do I declare usage-based annotation invoices on the EDF?
Each invoice is one row in Part 2B, with the amount billed and NRV. Where possible, invoice only accepted volume after the client's quality review, so the figures on the EDF match what the client will actually pay.
What if the client rejects work after I filed the EDF?
Issue a credit note, keep the EDF as filed, and write to your AD bank with the credit note explaining the reduced amount to be realised. Ask the bank how it wants the change recorded in EDPMS.
What SAC code should an AI training data company use?
It depends on how the work is framed and delivered. Labelling inside client platforms often points to IT services codes, while research support may point elsewhere. Confirm the code with your CA and use it consistently.
Are annotators working individually for foreign platforms covered?
RBI said on 7 October 2026 that individuals providing services abroad are not required to report and that FAQs would follow. Until then, individuals should check with their bank. Businesses should assume they are covered.
Bottom line
For annotation and AI training firms, invoice on accepted volume, document every quality credit and keep codes consistent. The EDF then simply mirrors your clean invoices. Prepare this month's declaration with the free EDF form generator.
This guide is general information for Indian businesses, not legal or tax advice. Rules and bank processes change — confirm specifics with your AD bank or chartered accountant.